an automated approach for epilepsy detection based on
TRANSCRIPT
Research ArticleAn Automated Approach for Epilepsy Detection Based onTunable Q-Wavelet and Firefly Feature Selection Algorithm
Ahmed I Sharaf 12 Mohamed Abu El-Soud13 and IbrahimM El-Henawy4
1Department of Computer Science Faculty of Computers and Information El-Mansoura University Egypt2Deanship of Scientific Research Umm Al-Qura University Mecca Saudi Arabia3Department of Computer Science University College of Umluj Tabuk University Saudi Arabia4Department of Computer Science Faculty of Computers and Information El-Zagazig University Egypt
Correspondence should be addressed to Ahmed I Sharaf ahmedsharaf84gmailcom
Received 20 April 2018 Revised 18 August 2018 Accepted 27 August 2018 Published 10 September 2018
Academic Editor A K Louis
Copyright copy 2018 Ahmed I Sharaf et alThis is an open access article distributed under the Creative CommonsAttribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
Detection of epileptic seizures using an electroencephalogram (EEG) signals is a challenging task that requires a high level ofskilled neurophysiologistsTherefore computer-aided detection provides an asset to the neurophysiologist in interpreting the EEGThis paper introduces a novel approach to recognize and classify the epileptic seizure and seizure-free EEG signals automaticallyby an intelligent computer-aided method Moreover the prediction of the preictal phase of the epilepsy is proposed to assist theneurophysiologist in the clinicThe proposedmethod presents two perspectives for the EEG signal processing to detect and classifythe seizures and seizure-free signalsThe first perspectives consider the EEG signal as a nonlinear time series A tunableQ-wavelet isapplied to decompose the signal into smaller segments called subbandsThen a chaotic statistical and power spectrum features setsare extracted from each subband The second perspectives process the EEG signal as an image hence the gray-level co-occurrencematrix is determined from the image to obtain the textures of contrast correlation energy and homogeneity Due to a large numberof features obtained a feature selection algorithm based on firefly optimization was applied The firefly optimization reduces theoriginal set of features and generates a reduced compact set A random forest classifier is trained for the classification and predictionof the seizures and seizure-free signals Afterward a dataset from the University of Bonn Germany is used for benchmarkingand evaluation The proposed approach provided a significant result compared with other recent work regarding accuracy recallspecificity F-measure and Matthewrsquos correlation coefficient
1 Introduction
Epilepsy is a chronic brain disease that affects people of allages According to the World Health Organization (WHO)approximately 65 million people suffer from this disorder [1]themajority of whom reside in developing countries and can-not obtain adequate medical treatment Epilepsy doubles ortriples the probability of sudden death when compared withthat for healthy people [2] Moreover epileptic patients sufferfrom social stigma and discrimination in their communitiesThis stigma has a negative impact upon the quality of lifeof patients and their families Therefore the investigation ofepilepsy detection techniques and antiepileptic drugs couldincrease the probability of those coping with this disease tolive healthily without social stigmas
Epilepsy is usually characterized by two or more unpro-voked seizures which affect the ictal person at any time Anelliptic seizure is defined as an excessive electrical dischargein an arbitrary portion of the brain This rapid dischargecauses a disturbance and abnormal behavior in the nervoussystem An adequate clinical tool used to recognize epilepticseizures is the EEG signal analysis as it measures the electro-physiological signals of the brain in real time and measuresbrain conditions efficiently [3] However EEG signal analysishas some limitations in detecting elliptic seizures because ofepilepsy behavior such as the following
(1) The occurrence of some seizures is not always becauseof the epilepsy disorder as approximately 10 ofhealthy people may suffer from one seizure in their
HindawiInternational Journal of Biomedical ImagingVolume 2018 Article ID 5812872 12 pageshttpsdoiorg10115520185812872
2 International Journal of Biomedical Imaging
lifetime These nonepileptic seizures are similar toepileptic seizures but they are not related to epilepsy[2] Hence the classification of both epileptic andnonepileptic seizures is further significant
(2) Although qualified professional neurologists canvisually detect epileptic seizures from an EEG datasheet it is still considered a time-consuming process
The diagnostics of epilepsy are usually performed by manualinspections of the EEG signals which not an easy task andrequires a highly skilled neurophysiologist Also the manualinspection of a long interval recording is a tedious andtime-consuming process Therefore an intelligent clinicalcomputer-aided design (CAD) tool that analyzes the EEGsignal and detects the epileptic seizure is required
Various case studies have reported the advantages ofusing automated methods to recognize epileptic seizuresfrom EEG signals Many techniques are commonly employedfor automated EEG analysis and epilepsy detection Mostof these techniques consist of two stages the first is con-cerned with feature extraction from the raw EEG signalthe other is dedicated to classifying the features [2] Thefeature extraction process is concerned with obtaining sig-nificant information from the raw EEG data as well as itcould be implemented in the time frequency and time-frequency domainsThe time domain and frequency domainare used for signal processing when the EEG is assumed tobe a stationary signal On the other hand when the EEGsignal is considered nonstationary [4 5] then the time-frequency domain is employed Case studies demonstratedthat the time-frequency domain is more suitable for EEGsignal analysis and could obtain significant results [2] Manyalgorithms have been proposed for elliptic seizure detectionwithin the time-frequency domain such as empirical modedecomposition (EMD) [6 7] and wavelet transformation [8ndash10] The EMD methods provided a leading trend to detectelliptic seizures from the EEG signal The EMD has beencombined with 2D and 3D phase space representation (PSR)features to identify elliptic seizures Then a least-squaressupport vector machine (LS-SVM) is used to perform theclassification process [11] A combination of different intrinsicmode functions (IMFs) is constructed as a set of featuresto utilize the classification problem [12] The EMD has alsobeen used to decompose an EEG signal into a collection ofsymmetric and band-limited signals Then a second-orderdifference plot (SODP) is applied to obtain an elliptical areaThe area under this shape with 95 confidence is usedas a selection measure fed to an artificial neural network(ANN) to determine the seizures and seizure-free signals [6]Although the EMDmethods proved their effectiveness thesemethods suffer from the mode-mixing problem which pro-duces intermediate signals and noise Local Binary Pattern(LBP) based methods represents a different approach of theepilepsy detection The work presented by [13] suggested afeature extraction based on one dimensional LBP to classifythe epileptic seizure seizure-free and the healthy classesfrom the EEG signal In [14] the researchers have imple-mented a technique based on the combination of the LBPand the Gabor filter of the EEG signals Then the k-nearest
neighbor classifier was used for the classification of epilepticseizures and seizure-free signals The wavelet transformationis usually employed with nonlinear measures to recognizeseizures and seizure-free patients from raw EEG signalsAn automatic epilepsy detection approach proposed by [15]used the discrete wavelet transformation (DWT) for signaldecomposition and generated a feature set using improvedcorrelation-based feature selection (ICFS)Then the randomforest classifier is applied for classification The DWT hasbeen usedwithmany nonlinear features and the effectivenessof this approach has been proved [16ndash23] Although wavelettransformation is an effectivemethod for EEG signal analysisthis transformation has some limitations [24] The selectionof an appropriate wavelet bias is vital in the time-frequencysignal analysis
A flexible wavelet transformation proposed by [25]namely tunableQ-wavelet transformation (TQWT) controlsthe transformation of a discrete time signal by an easilytunable variable called the Q-factor The TQWT solved theprimary limits of the wavelet filter banks by providing a tun-able Q-factor that controls the number of the oscillations ofthe wavelet transformation Moreover the TQWT decreasedthe search space of filter banks by providing three variablesonly for adjusting Also many researchers applied TQWT forphysiological signal analysis and proved its effectiveness [2122 26 27] However the after-mentioned methods provideda static set of features (eg statistical nonlinear and spectral)and did not discuss the adaptive behavior of these features asa dynamical system
In this paper an intelligent computer-aided design (CAD)tool that analyses the EEG signal and classifies the epilepticseizure and the seizure-free signal from the input EEG Thatprovides an asset to the neurophysiologist in interpreting theEEG and reduces the diagnostics timeThe proposedmethodis based on data fusion of a single-channel EEG signal and animage processing approach In the single-channel EEG signalthe EEG data are processed as a time-frequency time seriesThe signal is divided into smaller segments of data usingtunable Q-wavelet Some statistical features are extractedfrom this time series in the time domain and frequencydomain On the other hand an image processing techniqueextracts the significant texture from the medical imageThusthe gray-level co-occurrence matrix is applied to the imageand the contrast correlation energy and homogeneity areextracted The data fusion approach is used to combine thesefeatures of the input EEG signal and construct a large datasetfor each patient Because of a large number of the extractedfeatures a feature reduction algorithm is needed to reducethe processing time by obtaining a compact subset of featuresinstead of the original one Moreover the feature reductionalgorithm selects the relevant features removes redundantfeatures and discovers the dependency among these featuresTherefore the firefly algorithm is used to find the optimalsubset of features Consequently bootstraps are obtained byresampling the compact subset to train the random forestclassifier The final decision is obtained by performing a votefor each decision tree of the forest Hence the classificationof seizure and seizure-free is obtained A real-world datasetfrom the University of Bonn is used for benchmarking and
International Journal of Biomedical Imaging 3
validation of the proposed method A numerical experimenthas been implemented and a comparative study presenteda promising efficiency of the proposed system regarding theoverall accuracy sensitivity and specificity
The remainder of this manuscript is organized as followsthe preliminaries concepts were introduced in Section 2Section 3 introduced the combinational hybrid system ofthe epilepsy detection The experiment and discussion werepresented in Section 4 Lastly the paper was concluded inSection 5
2 Preliminary Knowledge
21 Tunable Q-Wavelet Transformation (TQWT) The tun-ability of the Q-factor provided a proficient method to adoptthe wavelet transformation [25] The TQWT have threeinputsQ-factor denoted by119876 which determines the numberof oscillations of the wavelet the number of the oversamplingrate which is denoted by 119903 andwhich determines the numberof the overlapping frequency responses and the number ofstages of decomposition denoted by 119869 For each decomposi-tion stage the target signal 119904[119899]with a sample rate of 119891119904 couldbe represented by low-pass and high-pass subbands withsampling frequencies of120572119891119904 and120573119891119904 respectively where120572 and120573 are the parameters of signal scaling The low-pass subbandis presented by low-pass filter 1198670(120596) and low-pass scaling119871119875119878(120572) Similarly the high-pass subband 1205961 is produced by1198671(120596) and 119867119875119878(120573) The low-pass and high-pass subbandsignals are formulated as follows
H1198880 (120596)=
0 120572120587 le |120596| le 120587120579(120596 + (120573 minus 1) 120587120572 + 120573 minus 1 ) (1 minus 120573) 120587 le |120596| lt 120572120587
1 |120596| lt (1 minus 120573) 120587
(1)
H1 (120596) =
0 |120596| lt (1 minus 120573) 120587120579( 120572120587 minus 120596120572 + 120573 minus 1) (1 minus 120573) 120587 le |120596| lt 1205721205871 120572120587 le |120596| le 120587
(2)
where 120579(120596) could be defined as follows
120579 (120596) = 05 (1 + cos (120596)) (2 minus cos (120596))05 |120596| le 120587 (3)
Both of 119903 and 119876 could be represented as filter-bank variables120572 and 120573 as follows
119903 = 1205731 minus 120572 119876 = 2 minus 120573120573 (4)
22 Feature Sets The feature sets used in this research aregrouped into four main groups which are statistical powerspectrum chaotic features and gray-level co-occurrencematrix (GLCM)The first group contains a set of five features
calculated from the time domain of the input signal Thisfeature set containsmean (120583) standard deviation (119878119879119863) vari-ance (var) Shannon entropy (119867) and approximate entropy(ApEn) The mathematical formulation of each feature isshown as follows [28ndash31]
120583 (119909) = 1119873119873sum119894=1
119909119894 (5)
119878119879119863 (119909) = radic 1119873 minus 1119873sum119894=1
(119909119894 minus 120583119909)2 (6)
var (119909) = 1119873 minus 1119873sum119894=1
1003816100381610038161003816119909119894 minus 12058310038161003816100381610038162 (7)
119867(119883) = 119873sum119894=1
119875 (119909119894) log (119875 (119909119894)) (8)
119860119901119864119899 (x) = 120601119898 (119903) minus 120601119898+1 (119903) (9)
The second set of features calculates the power spectrum ofthe input signal based on the frequency domain analysisThis feature set contains spectral centroid (119878119862) spectral speed(119878119878) spectral flatness (119878119865) spectral slope (119878119878119868) and spectralentropy (119875119878119864) where 119884(119902) denotes the for the discreteFourier transformation of the input signal 119891(119899) The math-ematical formulation of each feature is shown as follows [32]
119878119862 = sum119872minus1119902=0 119902 1003816100381610038161003816119884 (119902)1003816100381610038161003816sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816 (10)
119878119878 = sum119872minus1119902=0 (119902 minus 119878119862)2 1003816100381610038161003816119884 (119902)1003816100381610038161003816sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816 (11)
119878119865 = prod119872minus1119902=0 1003816100381610038161003816119884 (119902)10038161003816100381610038161119872(1119872)sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816 (12)
SSI = 119872sum119872minus1119902=0 119891119898 1003816100381610038161003816119884 (119902)1003816100381610038161003816 minus sum119872minus1119902=0 119891119898 sdot sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816119872sum119872minus1119902=0 1198912119898 minus (sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816)2 (13)
119875119878119864 = minus 119899sum119894=1
119884 (119902)sum119899119894=1 119884 (119902) ln(119884 (119902)sum119899119894=1 119884 (119902)) (14)
The third set of features contains chaotic measures to obtainthe dynamic behavior of the EEG signal This set includesHiguchirsquos fractal dimension (119867119865119863) Hurst exponent (119867119903)and Katz fractal exponent (119870119860119879119885) These features areformulated as follows [33ndash36]
119867119865119863 = ln( 119896119871 (119896)) (15)
119864[119877 (119899)119878 (119899) ] = 119888119899119867119903 (16)
119870119860119879119885 = log (119899)log (119899) + log (119889119871) (17)
4 International Journal of Biomedical Imaging
The final set of features consists of statistical measuresof an image represented as matrices called gray-level co-occurrence matrix (GLCM) where 119862(119894 119895) represents an entryin co-occurrence matrix and 119894 119895 = 0 1 2 119871 minus 1 where119871 is the number of gray levels in the image Those matricesrepresent the spatial dependencies between the gray levels ofimage reflecting the structure of the underlying texture Afterthe normalization of these matrices the contrast correlationenergy and homogeneity are computed as follows
119864119899119890119903119892119910 = 119871minus1sum119894=0
119871minus1sum119895=0
119862 (119894 119895)2 (18)
119862119900119899119905119903119886119904119905 = 119871minus1sum119894=0
119871minus1sum119895=0
(119894 minus 119895)2 119862 (119894 119895) (19)
119862119900119903119903119890119897119886119905119894119900119899 = 119871minus1sum119894=0
119871minus1sum119895=0
(119894 minus 120583119894) (119895 minus 120583119895) 119862 (119894 119895)120590119894120590119895 (20)
119871119900119888119886119897 ℎ119900119898119900119892119890119899119890119894119905119910 = 119871minus1sum119894=0
119871minus1sum119895=0
11 + (119894 minus 119895)2119862 (119894 119895) (21)
23 Firefly Optimization Algorithm The firefly algorithm isa swarm based stochastic search technique [37] The fireflyoptimization algorithm consists of a set of members calledfireflies each firefly represents a candidate solution Themost attractive firefly is considered to be the leader fireflythat leads the other candidates to the best region Theattractiveness is calculated based on the light intensity whichis usually determined by the objective fitness function Theattractiveness between two fireflies 119883119894 and 119883119895 is determinedas follows
120573 (119903119894119895) = 1205730119890minus1205741199032119894119895 (22)
119903119894119895 = radic 119863sum119889=1
(119909119894119889 minus 119909119895119889)2 (23)
where 119863 denotes the problem dimension such that 119863 =1 2 119889 119903119894119895 denotes the distance between 119883119894 and 119883119895Parameter 1205730 denotes the initial attractiveness at 119903 = 0 and120574 denotes the light absorption factor such that 120574 isin [0 1]Each firefly 119883119894 is compared with the other fireflies 119883119895 where119895 isin 1 2 119873 such that 119894 = 119895 and119873 denotes the count of thefireflies If firefly 119883119894 is better (brighter) than 119883119895 then firefly119883119895 moves towards 119883119894 with a step movement formulated asfollows
119883119894119889 (119905 + 1) = 119883119894119889 (119905) + 1205730119890minus1205741199032119894119895119894119895 (119883119895119889 (119905) minus 119883119894119889 (119905))+ 120572120598119894 (24)
where 120598119894 represents uniform a randomly distributed variablesuch that 120598119894 isin [minus05 05] and 120572 denotes the movement stepsuch that 120572 isin [0 1]
3 The Combinational Hybrid System ofEpilepsy Detection from EEG Signal
In this research a hybrid system was proposed to detectboth seizures and seizure-free conditions from a raw EEGsignal Although some investigations focused on the featureextraction level the proposed system was established basedon four main levels This system combined the data fusionapproach with firefly optimization and random forest The119879119876119882119879 was applied for EEG signal decomposition then thefeatures were constructed using a data fusion technique Dueto the large number of features obtained for each subband(119891119890119886119905119906119903119890119904119862119900119906119899119905 times 119869 times 119876) a feature reduction was applied toreduce the features and to obtain a compact set of featuresinstead of the original one The obtained compact set offeatures was fed to a random forest algorithm to obtainthe classification rules and hence used for training Aftertraining the classifier should be able to classify and estimatethe preictal phase The proposed system was divided intothe following four levels of processing and then described indetail as shown in Figure 1
(i) EEG decomposition using TQWT(ii) Feature extraction using data fusion based on single-
channel EEG signal and co-occurrence matrix(iii) Feature reduction using firefly optimization algo-
rithm(iv) Training of random forest classifier to detect the
seizures and seizure-free EEGs
31 EEG Decomposition Using TQWT The preprocessinglevel applies the 119879119876119882119879 decomposition to the input EEGsignal The 119879119876119882119879 converted the continuous EEG signalto discrete potions of data that could be handled moreeffectively This wavelet transformation is used because of itseffectiveness in signal decomposition and its tunability Theobtained subbands using the 119879119876119882119879 provided a significantdifference between the seizure-free and the epileptic seizureof the EEG signals as shown in Figure 2 The subfiguresdenoted by (a) (b) (c) and (d) visualize the histogram of thefirst second third and fourth subbands of the seizure-freeclass The remaining subfigures denoted by (e) (f) (g) and(h) represent the histogram of the epileptic seizure class forthe same subbands The values of the extracted subbands ofthe second class are about ten times stronger than the firstclass that prove the efficiency of this decomposition
32 Feature Extraction Using Data Fusion Based on a Single-Channel EEG and a Co-Occurrence Matrix In the firstperspective the EEG signal was described as a nonstationarytime series After the decomposition of the EEG signal usingthe TQWT a feature extraction process was performed toobtain significant characteristics fromeachTQWTsubbandsThe extracted features were categorized into three maingroups The first group determines the statistical character-istics in the time domain The mean standard deviationvariance Shannon entropy and approximate entropy werecalculated in the first group as formulated from (5) to (9)
International Journal of Biomedical Imaging 5
Image Extract
chaotic features Apply TQWT for the signal decomposition Train the random forest
classifier
EEG Signal
Extractstatistical features Feature reduction using
firefly algorithm End Constrcut feature space using data
fusion
Classifiy and estimate seizures using the trained
model
Single EEG Channel
Extract power spectrum
features
Extract Co-occurrence
matrix
Input Level 1 Level 2 Level 3 Level 4
Figure 1 Block diagram of the combinational hybrid system of epilepsy detection from EEG signal with 4 levels of processing
2500
2000
1500
1000
500
0
occu
ranc
e
minus20 0 20
V
(a)
0 20 40minus20minus40
V
0
500
1000
1500
2000
2500
3000
occu
ranc
e
(b)
0
200
400
600
800
1000
1200
1400
0 50 100minus50
V
occu
ranc
e
(c)
0
200
400
600
800
1000
1200
0 200 400minus200
V
occu
ranc
e
(d)
0
500
1000
1500
2000
2500
0 200minus200
V
occu
ranc
e
(e)
2500
2000
1500
1000
500
00 200 400minus200minus400
V
occu
ranc
e
(f)
0 500minus500
V
0
200
400
600
800
1000
1200
occu
ranc
e
(g)
0
200
400
600
800
1000
1200
0 500 1000minus500minus1000
V
occu
ranc
e
(h)
Figure 2 Illustration of seizure-free and epileptic seizures of subbands obtained from TQWT with Q = 1 r = 3 J = 3 Figures (a) (b) (c)and (d) represent the first second third and fourth subbands obtained from seizure-free signals Figures (e) (f) (g) and (h) represent thefirst second third and fourth subbands obtained from the epileptic seizures
This group indicates some statistical information obtainedfrom the time domain of TQWT subband The secondgroup of features determines a power spectrum analysisof obtained subbands The discrete Fourier transformations(DFT) were applied to convert these subbands into a fre-quency domain Then the power spectrum features wereextractedThe second group consists of the spectral centroidspectral speed spectral flatness spectral slope and spectralentropy as formulated from (10) to (14) The power spectrumanalysis represents an effectivemethod to study the frequencybehavior of the signal The last group of features performs achaotic analysis of each subband Because of the nonlinearityof the EEG signals a nonlinear analysis is required One ofthe best analyses used for this issue is the chaotic analysisIn this analysis the Higuchi fractal dimension (HFD) Hurst
exponent and Katz fractal exponent were computed for eachsubband as formulated from (15) to (18)
In the second perspective the input EEG signal wasconverted to a gray image Then a co-occurrence matrix wascomputed to obtain the gray levels of the image Then thetextures of contrast correlation energy and homogeneitywere calculated from this matrix to represent the statisticalmeasures of the image as formulated from (19) to (21) Finallyafter computing the feature space a data fusion was appliedto merge all of these features and create a single dataset
33 Feature Reduction Using Firefly Optimization AlgorithmIn this section a feature reduction algorithm based onthe firefly algorithm is proposed [37 38] This algorithmimplements a chaotic movement simulated annealing (SA)
6 International Journal of Biomedical Imaging
(1) InputThe features matrix 119865119898(2) OutputThe optimal solution 119892119887119890119904119905(3) Initialize the firefly swam using a logistic chaotic map(4) Evaluate each firefly 119892119894 using the fitness function f (x)(5) Select the best firefly 119892119887119890119904119905 and the worst one 119892119908119900119903119904119905(6) while termination condition are not reached do(7) Declare an alternative leader firefly as 119878119887119890119904119905 with a competitive fitness and located in different region(8) Obtain an offspring solution 119892119887119890119904119905 using 119878119860(9) for all (firefly 119894 and 119891119894 = 119892119908119900119904119903119905) do(10) for all (firefly 119895 and 119891119895 = 119892119908119900119904119903119905) do(11) if (119868119895 gt 119868119894) then(12) Enhance firefly 119895 using Equation (25) to obtain better offspring candidate firefly denoted 1199091015840119895(13) Replace firefly 119895 with the offspring 1199091015840119895(14) Move the firefly 119895 towards the neighboring and global optimal solutions using Equation (26)(15) end if(16) end for(17) end for(18) Update the worst solution 119892119908119900119903119904119905(19) if 119891(1198921015840119887119890119904119905) gt 119891(119892119887119890119904119905) then(20) 119892119887119890119904119905 larr997888 1198921015840119887119890119904119905(21) end if(22) Rank all fireflies and update the best 119892119887119890119904119905 and worst 119892119908119900119903119904119905 solutions(23) end while(24) return 119892119887119890119904119905Algorithm 1 The pseudocode of the proposed feature reduction algorithm based on firefly optimization and SA
to produce efficient offspring candidates andmemory aware-ness of the best and worst solutions to improve the searchdiversity and prevent local optima The firefly population israndomly initialized using a chaotic logistic map to ensurethe diversely of the candidates and the randomness of eachfirefly Afterwards the fitness function is computed for eachcandidate and identifies both the best and worst solutions as119892119887119890119904119905 and 119892119908119900119904119903119905 respectively For each iteration an alternativecandidate is declared as 119878119887119890119904119905 with a competitive fitnessand located in a different region Both of the best and thealternative candidate sets are used to lead weak solutions toreach the optimal region and prevent local optima problemThe mean of the leader firefly and the alternative one isenhanced using SA algorithm to obtain a better solution1198921015840119887119890119904119905 The improved local and global solutions are used toguide the low lightness fireflies to move towards that strongerlightness The algorithm is repeated until the maximumnumber of iterations is reached or a termination criterion isachieved The behavior of the proposed algorithm is deter-mined by some properties namely the objective function theattractiveness movement step and population diversity Theproposed algorithm is demonstrated in Algorithm 1
TheObjective FunctionThis function is used to evaluate eachcandidate in the algorithm and defined as follows
119891 (119909) = 1199081 times 119886119888119888119906119903119886119888119910 (119909) + 1199082number of features (25)
where 1199081 and 1199082 represent the weights of the classifi-cation accuracy and the number of the selected features
respectively The values of 1199081 and 1199082 are set to 09 and 01respectively as a recommended by [38]
The Attractiveness Movement Step The proposed algorithmused a chaotic logistic map to initialize the firefly populationand hence increases the diversity and avoid local optimaAfter obtaining the global best solution 119892119887119890119904119905 an alternativeleader firefly 119878119887119890119904119905 is declared with a competitive fitness butlocated in a different region Since both leaders are morelikely to discover distinctive search regions this strategyreduces the probability of being trapped in the local optimaIn addition the optimal offspring of themean positions of thetwo leaders and the neighboring brighter candidates are usedto lead the search process and guide the solutions with lowerlight intensity to move towards the optimal region
119909119894 = 119909119894 + 1205730119862119896 (1199091015840119895 minus 119909119894) + 119862119896120576 (1198921015840119887119890119904119905 minus 119909119894) + 1205721015840times sign [119903119886119899119889 minus 05] (26)
1199091015840119895 = 119909119895 + 1205901 (27)
1198921015840119887119890119904119905 = 119898119890119886119899 (119892119887119890119904119905 + 119878119887119890119904119905) + 1205902 (28)
where 1199091015840119895 denotes the offspring candidate with a brighterneighboring solution119909119895 is defined by the SA as shown in (26)and 1198921015840119887119890119904119905 represents the fitter offspring solutions of the meanof the leader firefly and the alternative one as formulated in(27) It worth mentioning that the values of 1205901 and 1205902 aretwo random variables set using the Gaussian distributionThe movement step of the firefly is determined as shown in
International Journal of Biomedical Imaging 7
(1) Input 119894119905119890119903119886119905119894119900119899119904119898119886119909 119892119898119890119886119899 119879119898119886119909(2) OutputThe optimal offspring 119878119887119890119904119905(3) 119878119888 = 119862119903119890119886119905119890119868119899119894119905119878119900119906119897119905119894119900119899119904(119892119898119890119886119899)(4) 119878119887119890119904119905 = 119878119888(5) while (119894 le 119894119905119890119903119886119905119894119900119899119904 119898119886119909) do(6) 119878119894 = 119862119903119890119886119905119890119873119890119894119892ℎ119887119900119903119878119900119897119906119905119894119900119899(119878119888)(7) 119879119888 = 119862119886119897119888119906119897119886119905119890119879119890119898119901119890119903119886119905119906119903119890(119894 119879119898119886119909)(8) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119888)) then(9) 119878119888 = 119878119894(10) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119887119890119904119905)) then(11) 119878119887119890119904119905 = 119878119894(12) end if(13) else if (exp((119888119900119904119905(119878119887119890119904119905) minus 119888119900119904119905(119878119894))119879119888) ge 120590) then(14) 119878119888 = 119878119894(15) end if(16) 119894 = 119894 + 1(17) end while(18) return 119878119887119890119904119905Algorithm 2 The pseudocode of the generation of offspring solutions using SA
(28) where 119862119870 represents the chaotic map variable in themovement step and 120576 denotes the randomized vector definedin the traditional firefly algorithm Parameter 1205721015840 denotes anadaptive step initialized to 05 to control the diversity of thesearch process
Offspring Generation Using SAThe proposed algorithm usedSA for generating better candidates to enhance the searchprocess as much as possible as shown in Algorithm 2 TheSA accepts both of the best solution 119892119887119890119904119905 and the alternativesolution 119878119887119890119904119905 as main inputs then the traditional SA isapplied The better solution generated is accepted by defaultaccording to the SA heuristics On the other hand theweaker solution should be accepted with specific probabilityas shown in (29) where Δ119891 denotes the difference of thefitness (energy) between to candidates and 119879119888 denotes thecurrent temperature A simple linear cooling mechanism isused to control the value of the temperature
119901119909119895 = exp(Δ119891119879119888 ) (29)
Population Diversity In each iteration the worst solution isdetected after the ranking process as shown in Algorithm 1The remaining solutions are guided by the average positionobtained from (30) where 120590 denotes a random value obtainedby Gaussian map
119909119908119900119903119904119905119895 = 119892119887119890119904119905 + 1198781198871198901199041199052 + 120590 (119909119908119900119903119904119905119895 119892119887119890119904119905 + 1198781198871198901199041199052 ) (30)
34 Classification and Learning The random forest (RF) isa successful ensemble approach used in supervised machinelearning to solve classification or regression problems [39]It consists of a collection of decision trees that could actas a single classifier with multiple classification methodsor a method that has several variables Several subsets ofthe training data are supplied to each tree to achieve the
most stable tree classification that results in a generalizedexperience of the classifier The original dataset is dividedinto two parts The first part is used to train each tree bybootstrapping technique The other part is used to evaluatethe accuracy of the classification Each tree is allowed to reachthe maximum depth without tree pruning to obtain a highvariance classifier The splitting process remains until onlyone instance of a single class is dropped from any leaf nodeor a predefined termination condition is achieved Whenthe forest is established the number of subsets remained asa constant The obtained route of traversal from the rootnode to the leaf node is applied to the new instances orthe unlabeled instance for classification The final decisionfor classifying a new instance is provided by determiningeach class that has the most votes from every decision treeThe random forest performs slightly better when comparedwith other classifiers such as discriminant analysis SVM andartificial neural network (ANN) [15 39]
35 Performance Evaluation Various performance formulaswere used to evaluate the effectiveness of a classifierThe sen-sitivity (SEN) or recall specificity (SPEC) accuracy (ACC)F-measure Matthewrsquos correlation coefficient (MCC) andreceiver operating characteristics (ROC) are used to evaluatethe efficiency of the random forest classifier [40ndash42] Theseparameters are defined as follows
119878119864119873 = 119879119875119879119875 + 119865119873 times 100 (31)
119878119875119864119862 = 119879119873119879119873 + 119865119875 times 100 (32)
119860119862119862 = 119879119875 + 119879119873119879119875 + 119879119873 + 119865119875 + 119865119873 times 100 (33)
119865 minus 119898119890119886119904119906119903119890 = 21198791198752119879119875 + 119865119875 + 119865119873 times 100 (34)
8 International Journal of Biomedical Imaging
119872119862119862= (119879119875 times 119879119873) minus (119865119875 times 119865119873)radic(119879119875 + 119865119873) (119879119875 + 119865119875) (119879119873 + 119865119873) (119879119873 + 119865119875)times 100
(35)
119879119875 and 119879119873 represent the total number of an epileptic seizureand seizure-free signals classified correctly respectively Sim-ilarly 119865119875 and 119865119873 represent the total number of epilepticseizures and seizure-free signals classified incorrectly respec-tively Cross-validation has also been used to ensure theclassifier reliability and effectiveness The original dataset issplit into 119896 folds (subsets) for both training and testing Inthis strategy 119896 minus 1 folds were selected randomly to train theclassifier and the remaining folds are used to testing Theoverall performance is calculated as the average of each foldIn this work the experiment is repeated 10 times with tenfoldcross-validation
4 Results and Discussion
The benchmark dataset used in this investigation wasacquired by the University of Bonn [28]The dataset containsthree different categories ie preictal healthy and ictalrecorded using a single channel for a 236 s duration Bothnormal and preictal conditions were collected from 200 casestudies and 100 for the ictal state The normal condition isacquired from five healthy volunteers using the international10ndash20 system standard with each volunteer in a relaxed-awake state with eyes open and closed (100 cases per each set)[28] The ictal data were collected from five patients duringtheir epileptic seizures The preictal represents the EEG datacollected from the same five patients with no seizures It isworth mentioning that all EEG signals were acquired using a128 channel amplifier with sampling rate equal 17361Hz [28]Finally a bandpass filter with 05340Hz sim12 dBoctave wasapplied as a filter
The proposed approach for automatic detection of epilep-tic seizures and seizure-free patients was implemented usingMATLAB software A TQWT comparison between theseizure-free and the epileptic seizures patients are shown inFigure 3 with 119876 = 1 119903 = 3 and 119869 = 3 It can be observedthat both of the amplitudes and the frequency of the epilepticseizure are much higher than the healthy one Moreover theoscillatory behavior of the epileptic patient is higher than thehealthy one The value of the parameter 119903 is set to three toprevent any excessive ringing of the wavelet as suggested by[22] MATLAB for the TQWT toolbox is available for publicaccess at httpeewebpolyeduiselesniTQWT
After the construction of the dataset the feature reduc-tion was applied using the firefly algorithm to remove theredundant and irrelevant features The number of the datasegments was determined by the parameters 119876 119903 and 119869 Bytuning these parameters the number of the data segmentswas varied and thus the training process was adopted Thetrial and error approach was used to set the value of theseparameters The experiment was implemented on variousvalues of J to obtain the best level of decomposition The
performance measures were calculated for each level ofdecomposition As shown in Figure 4 the best value of thevariable 119869 was from two to three Moreover the best value ofthe parameter119876was found to be one as shown in Figure 5 Allthe variables remained constant while changing the Q-factorto obtain the best value
Then the firefly algorithm with a population size of 20fireflies mutation probability of 001 and light absorptionequal to 01 was applied to reduce the feature setThe result ofthe compact set of features is obtained after 100 iterations andshown in Figure 6 are the number of features obtained by thefirefly algorithm and their corresponding accuracy precisionspecificity and the recall In the last step the feature set isfed to a random forest classifier to obtain the seizure-free andseizure conditionsThe random forest classifier obtained 98accuracy 97 precision 97 specificity 98 recall 98 F-measure and 95 MCC at the third level of decompositionand the Q-factor equals 1
The firefly algorithm reduced the search space into threefeatures These features could replace the original datasetwhich minimizes the processing time The compact set offeature contains the 119878119879119863 119860119901119864119899 and the 119870119860119879119885 the classi-fication rules are based on these features The classificationrules obtained by the proposed system are shown in Figure 7where the decision tree consists of 4 leaves and 7 nodes andthe final decision represents the seizure-free and the epilepticseizure denoted by 01 respectively
A comparative study of the proposed hybrid epilepsydetection approach and other existing classification systemshas been performed in terms of the total accuracy A novelmethod based on the EMDs was proposed to detect theepileptic seizures of epilepsy This method used the Hilberttransformation of IMFs obtained by EMD process thatprovided an analytic signal representation of IMFs [12] Theclassification rules obtained by this method achieved anaccuracy of 90 The usage of frequency domain featuresand Burgrsquos method obtained 9311 accuracy with SVMclassifier [43] Nonlinear features have been used with aGaussianmixturemodel classifier and achieved 95accuracy[44] A decision tree classifier is used with energy fractaldimension and sample entropy and provided 957 [45]A combination of ApEn and the Hurst exponent has beenused to detect the diagnostics of epilepsy and produced 965accuracy with SVM classifier and ANN [46] In [47] aneigensystem based method was proposed cooperated withMultiple Layer Perceptron to classify the epileptic seizureshealthy and the seizure-free This approach provided anaverage accuracy of 975 The EMD methods for epilepsydetection achieved 9775 accuracy [6]The Kraskov entropyis also combinedwith the SVMclassifier and provided 9775[22] An automated diagnostics system based on a set ofentropies and fuzzy Sugeno classifier (FSC) achieved accu-racy up to 98 [19] The work presented by [48] developeda method for the epilepsy detection using the EMD Thegenerated IMFs using the EMD were represented as a set ofamplitude and frequency modulated (AMndashFM) signals Thetwo bandwidths namely amplitude modulation bandwidthand frequency modulation bandwidth calculated from theanalytic IMFs have been fed to LS-SVM for classifying
International Journal of Biomedical Imaging 9
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(a) EEG Sample for patient with seizure-free
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(b) EEG Sample for patient with seizures
Figure 3 TQWT decomposition of seizure-free and seizure EEG signals with 119876 = 1 119903 = 3 119895 = 3
1 2 3 4 5 10LEVELS OF DECMPOSISTION (J)
ACCPrecision
SPECRecall
000
020
040
060
080
100
PERC
ENTA
GE
Figure 4 Performance measures of the proposed approach withvaried levels of decomposition
08082084086088
09092094096098
1
1 2 3 5 7 9 10
PERC
ENTA
GE
Q
ACCPrecision
SPECRecall
Figure 5 Performance measures of the proposed approach withvaried levels of decomposition
seizure and nonseizure EEG signals This method achieved9818 average accuracy The LBP-based methods have beencombined with Gabor filter for texture extraction from theEEG signal A k-nearest neighbor classifier was applied andobtained an accuracy of 983 [14] The proposed methodconfirmed its superiority in the total accuracy compared tothe other systems The results which prove the superiority ofthe proposed method compared to the other existed systemsis demonstrated in Figure 8
ACCPrecision
SPECRecall
09091092093094095096097098099
1
1 2 3 4 5PE
RCEN
TAG
ENUMBER OF FEATURES
Figure 6 Performance measures of the original and compactfeatures set
STD
lt= 2873 gt 2873
0
0
ApEn
lt= 016 gt 016
Katz 1
1
lt= 146 gt 146
Figure 7 The classification rules obtained from the proposedsystem to classify the seizure-free (0) and the epileptic seizure (1)signals
Once the system detects the preictal phase the clinicreceives a notification about that patientThemain advantageof the hybrid system from the clinical point of view could besummarized as follows
(i) Classification and detecting of the epileptic seizuresand seizure-free signals from the EEG signal automat-ically
10 International Journal of Biomedical Imaging
90
931195 957 965
975 9775 9775 98 9813 983 99
Baja
j et a
l 20
13
Faus
t et a
l 20
10
Acha
rya e
t al
2009
Mar
tis et
al 2
013
Yuan
et al
201
1
Nag
hsh-
Nilc
hi et
al 2
010
Pach
ori e
t al
2014
Patid
ar et
al 2
017
Acha
rya e
t al
2012
Baja
j et a
l 20
12
Kum
ar et
al 2
015
e p
ropo
sed
met
hod
PROPOSED SYSTESMS USED TO FOR THE COMPARSION
8486889092949698
100
ACCU
RACY
IN P
ERCE
NTA
GE
()
Figure 8 The total accuracy () of various methods employed to detect the disorder of the epilepsy
(ii) The detection of the preictal phase from studying thehealthy and ictal phase of each patient
(iii) The detection of the preictal phase that provided theability to send warning alert to the physicians toprepare the medical assessment for the patient
(iv) The proposed method being robust and reliable as itsperformance was benchmarked using 10-fold cross-validation
(v) The dynamic behavior obtained from the fireflyalgorithm which made the system adaptive to manyfeatures according to each case study
(vi) A few sets of parameters required to analyze the EEGsignal typical three variables (119876 119903 119869)
The limitations of this research were summarized as follows
(i) The limited number of the studied subjects (typically100 per class)
(ii) The diagnosis process that may be reduced because ofdepending on some additional software prerequisites
5 Conclusion
In this paper an automated intelligent CAD tool has beenproposed to classify and detect epileptic seizures and seizure-free EEG signals This method provided an EEG signalanalysis using a hybrid data fusion method The data fusionmethod combined the collected features from two differentperspectives In the first perspective the EEG signal wasconsidered as an image Then the image was converted toa gray image and the GLCM was obtained to extract thetextures of the image such as contrast correlation power andhomogeneity In the second perspective the EEG signal wasdivided into smaller segments using the TQWT to extractthe time and frequency features A set of statistical nonlinear(chaotic) and power spectrum features were obtained from
each segment After the dataset was constructed a featurereduction algorithmbased on firefly optimizationwas used toreduce the irrelevant features and remove redundancy Thenan RFwas trained to classify and predict the epileptic seizuresand seizure-free EEG signals from the dataset The experi-mental results showed that the proposed method achieveda satisfactory degree of 99 accuracy 97 precision 97specificity 98 recall 98 F-measure and 95 MCC at thethird level of decomposition
Data Availability
The dataset used to support the findings of this research wasincluded in the article and can be found as a citation to ldquoRGAndrzejak K Lehnertz F Mormann C Rieke P DavidCE Elger Indications ofNonlinearDeterministic and Finite-Dimensional Structures in Time Series of Brain ElectricalActivity Dependence on Recording Region and Brain StatePhysical Review E 64 (2001) 061907rdquo
Conflicts of Interest
The authors declare that they have no conflicts of interest
References
[1] Epilepsy Fact Sheet 2018 httpwwwwhointmediacentrefactsheetsfs999en
[2] U R Acharya S V Sree G Swapna R J Martis and J S SurildquoAutomated EEG analysis of epilepsy a reviewrdquo Knowledge-Based Systems vol 45 pp 147ndash165 2013
[3] A Puce and M S Hamalainen ldquoA review of issues relatedto data acquisition and analysis in EEGMEG studiesrdquo BrainSciences vol 7 no 6 2017
[4] V Joshi R B Pachori and A Vijesh ldquoClassification of ictaland seizure-free EEG signals using fractional linear predictionrdquoBiomedical Signal Processing and Control vol 9 pp 1ndash5 2014
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
2 International Journal of Biomedical Imaging
lifetime These nonepileptic seizures are similar toepileptic seizures but they are not related to epilepsy[2] Hence the classification of both epileptic andnonepileptic seizures is further significant
(2) Although qualified professional neurologists canvisually detect epileptic seizures from an EEG datasheet it is still considered a time-consuming process
The diagnostics of epilepsy are usually performed by manualinspections of the EEG signals which not an easy task andrequires a highly skilled neurophysiologist Also the manualinspection of a long interval recording is a tedious andtime-consuming process Therefore an intelligent clinicalcomputer-aided design (CAD) tool that analyzes the EEGsignal and detects the epileptic seizure is required
Various case studies have reported the advantages ofusing automated methods to recognize epileptic seizuresfrom EEG signals Many techniques are commonly employedfor automated EEG analysis and epilepsy detection Mostof these techniques consist of two stages the first is con-cerned with feature extraction from the raw EEG signalthe other is dedicated to classifying the features [2] Thefeature extraction process is concerned with obtaining sig-nificant information from the raw EEG data as well as itcould be implemented in the time frequency and time-frequency domainsThe time domain and frequency domainare used for signal processing when the EEG is assumed tobe a stationary signal On the other hand when the EEGsignal is considered nonstationary [4 5] then the time-frequency domain is employed Case studies demonstratedthat the time-frequency domain is more suitable for EEGsignal analysis and could obtain significant results [2] Manyalgorithms have been proposed for elliptic seizure detectionwithin the time-frequency domain such as empirical modedecomposition (EMD) [6 7] and wavelet transformation [8ndash10] The EMD methods provided a leading trend to detectelliptic seizures from the EEG signal The EMD has beencombined with 2D and 3D phase space representation (PSR)features to identify elliptic seizures Then a least-squaressupport vector machine (LS-SVM) is used to perform theclassification process [11] A combination of different intrinsicmode functions (IMFs) is constructed as a set of featuresto utilize the classification problem [12] The EMD has alsobeen used to decompose an EEG signal into a collection ofsymmetric and band-limited signals Then a second-orderdifference plot (SODP) is applied to obtain an elliptical areaThe area under this shape with 95 confidence is usedas a selection measure fed to an artificial neural network(ANN) to determine the seizures and seizure-free signals [6]Although the EMDmethods proved their effectiveness thesemethods suffer from the mode-mixing problem which pro-duces intermediate signals and noise Local Binary Pattern(LBP) based methods represents a different approach of theepilepsy detection The work presented by [13] suggested afeature extraction based on one dimensional LBP to classifythe epileptic seizure seizure-free and the healthy classesfrom the EEG signal In [14] the researchers have imple-mented a technique based on the combination of the LBPand the Gabor filter of the EEG signals Then the k-nearest
neighbor classifier was used for the classification of epilepticseizures and seizure-free signals The wavelet transformationis usually employed with nonlinear measures to recognizeseizures and seizure-free patients from raw EEG signalsAn automatic epilepsy detection approach proposed by [15]used the discrete wavelet transformation (DWT) for signaldecomposition and generated a feature set using improvedcorrelation-based feature selection (ICFS)Then the randomforest classifier is applied for classification The DWT hasbeen usedwithmany nonlinear features and the effectivenessof this approach has been proved [16ndash23] Although wavelettransformation is an effectivemethod for EEG signal analysisthis transformation has some limitations [24] The selectionof an appropriate wavelet bias is vital in the time-frequencysignal analysis
A flexible wavelet transformation proposed by [25]namely tunableQ-wavelet transformation (TQWT) controlsthe transformation of a discrete time signal by an easilytunable variable called the Q-factor The TQWT solved theprimary limits of the wavelet filter banks by providing a tun-able Q-factor that controls the number of the oscillations ofthe wavelet transformation Moreover the TQWT decreasedthe search space of filter banks by providing three variablesonly for adjusting Also many researchers applied TQWT forphysiological signal analysis and proved its effectiveness [2122 26 27] However the after-mentioned methods provideda static set of features (eg statistical nonlinear and spectral)and did not discuss the adaptive behavior of these features asa dynamical system
In this paper an intelligent computer-aided design (CAD)tool that analyses the EEG signal and classifies the epilepticseizure and the seizure-free signal from the input EEG Thatprovides an asset to the neurophysiologist in interpreting theEEG and reduces the diagnostics timeThe proposedmethodis based on data fusion of a single-channel EEG signal and animage processing approach In the single-channel EEG signalthe EEG data are processed as a time-frequency time seriesThe signal is divided into smaller segments of data usingtunable Q-wavelet Some statistical features are extractedfrom this time series in the time domain and frequencydomain On the other hand an image processing techniqueextracts the significant texture from the medical imageThusthe gray-level co-occurrence matrix is applied to the imageand the contrast correlation energy and homogeneity areextracted The data fusion approach is used to combine thesefeatures of the input EEG signal and construct a large datasetfor each patient Because of a large number of the extractedfeatures a feature reduction algorithm is needed to reducethe processing time by obtaining a compact subset of featuresinstead of the original one Moreover the feature reductionalgorithm selects the relevant features removes redundantfeatures and discovers the dependency among these featuresTherefore the firefly algorithm is used to find the optimalsubset of features Consequently bootstraps are obtained byresampling the compact subset to train the random forestclassifier The final decision is obtained by performing a votefor each decision tree of the forest Hence the classificationof seizure and seizure-free is obtained A real-world datasetfrom the University of Bonn is used for benchmarking and
International Journal of Biomedical Imaging 3
validation of the proposed method A numerical experimenthas been implemented and a comparative study presenteda promising efficiency of the proposed system regarding theoverall accuracy sensitivity and specificity
The remainder of this manuscript is organized as followsthe preliminaries concepts were introduced in Section 2Section 3 introduced the combinational hybrid system ofthe epilepsy detection The experiment and discussion werepresented in Section 4 Lastly the paper was concluded inSection 5
2 Preliminary Knowledge
21 Tunable Q-Wavelet Transformation (TQWT) The tun-ability of the Q-factor provided a proficient method to adoptthe wavelet transformation [25] The TQWT have threeinputsQ-factor denoted by119876 which determines the numberof oscillations of the wavelet the number of the oversamplingrate which is denoted by 119903 andwhich determines the numberof the overlapping frequency responses and the number ofstages of decomposition denoted by 119869 For each decomposi-tion stage the target signal 119904[119899]with a sample rate of 119891119904 couldbe represented by low-pass and high-pass subbands withsampling frequencies of120572119891119904 and120573119891119904 respectively where120572 and120573 are the parameters of signal scaling The low-pass subbandis presented by low-pass filter 1198670(120596) and low-pass scaling119871119875119878(120572) Similarly the high-pass subband 1205961 is produced by1198671(120596) and 119867119875119878(120573) The low-pass and high-pass subbandsignals are formulated as follows
H1198880 (120596)=
0 120572120587 le |120596| le 120587120579(120596 + (120573 minus 1) 120587120572 + 120573 minus 1 ) (1 minus 120573) 120587 le |120596| lt 120572120587
1 |120596| lt (1 minus 120573) 120587
(1)
H1 (120596) =
0 |120596| lt (1 minus 120573) 120587120579( 120572120587 minus 120596120572 + 120573 minus 1) (1 minus 120573) 120587 le |120596| lt 1205721205871 120572120587 le |120596| le 120587
(2)
where 120579(120596) could be defined as follows
120579 (120596) = 05 (1 + cos (120596)) (2 minus cos (120596))05 |120596| le 120587 (3)
Both of 119903 and 119876 could be represented as filter-bank variables120572 and 120573 as follows
119903 = 1205731 minus 120572 119876 = 2 minus 120573120573 (4)
22 Feature Sets The feature sets used in this research aregrouped into four main groups which are statistical powerspectrum chaotic features and gray-level co-occurrencematrix (GLCM)The first group contains a set of five features
calculated from the time domain of the input signal Thisfeature set containsmean (120583) standard deviation (119878119879119863) vari-ance (var) Shannon entropy (119867) and approximate entropy(ApEn) The mathematical formulation of each feature isshown as follows [28ndash31]
120583 (119909) = 1119873119873sum119894=1
119909119894 (5)
119878119879119863 (119909) = radic 1119873 minus 1119873sum119894=1
(119909119894 minus 120583119909)2 (6)
var (119909) = 1119873 minus 1119873sum119894=1
1003816100381610038161003816119909119894 minus 12058310038161003816100381610038162 (7)
119867(119883) = 119873sum119894=1
119875 (119909119894) log (119875 (119909119894)) (8)
119860119901119864119899 (x) = 120601119898 (119903) minus 120601119898+1 (119903) (9)
The second set of features calculates the power spectrum ofthe input signal based on the frequency domain analysisThis feature set contains spectral centroid (119878119862) spectral speed(119878119878) spectral flatness (119878119865) spectral slope (119878119878119868) and spectralentropy (119875119878119864) where 119884(119902) denotes the for the discreteFourier transformation of the input signal 119891(119899) The math-ematical formulation of each feature is shown as follows [32]
119878119862 = sum119872minus1119902=0 119902 1003816100381610038161003816119884 (119902)1003816100381610038161003816sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816 (10)
119878119878 = sum119872minus1119902=0 (119902 minus 119878119862)2 1003816100381610038161003816119884 (119902)1003816100381610038161003816sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816 (11)
119878119865 = prod119872minus1119902=0 1003816100381610038161003816119884 (119902)10038161003816100381610038161119872(1119872)sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816 (12)
SSI = 119872sum119872minus1119902=0 119891119898 1003816100381610038161003816119884 (119902)1003816100381610038161003816 minus sum119872minus1119902=0 119891119898 sdot sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816119872sum119872minus1119902=0 1198912119898 minus (sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816)2 (13)
119875119878119864 = minus 119899sum119894=1
119884 (119902)sum119899119894=1 119884 (119902) ln(119884 (119902)sum119899119894=1 119884 (119902)) (14)
The third set of features contains chaotic measures to obtainthe dynamic behavior of the EEG signal This set includesHiguchirsquos fractal dimension (119867119865119863) Hurst exponent (119867119903)and Katz fractal exponent (119870119860119879119885) These features areformulated as follows [33ndash36]
119867119865119863 = ln( 119896119871 (119896)) (15)
119864[119877 (119899)119878 (119899) ] = 119888119899119867119903 (16)
119870119860119879119885 = log (119899)log (119899) + log (119889119871) (17)
4 International Journal of Biomedical Imaging
The final set of features consists of statistical measuresof an image represented as matrices called gray-level co-occurrence matrix (GLCM) where 119862(119894 119895) represents an entryin co-occurrence matrix and 119894 119895 = 0 1 2 119871 minus 1 where119871 is the number of gray levels in the image Those matricesrepresent the spatial dependencies between the gray levels ofimage reflecting the structure of the underlying texture Afterthe normalization of these matrices the contrast correlationenergy and homogeneity are computed as follows
119864119899119890119903119892119910 = 119871minus1sum119894=0
119871minus1sum119895=0
119862 (119894 119895)2 (18)
119862119900119899119905119903119886119904119905 = 119871minus1sum119894=0
119871minus1sum119895=0
(119894 minus 119895)2 119862 (119894 119895) (19)
119862119900119903119903119890119897119886119905119894119900119899 = 119871minus1sum119894=0
119871minus1sum119895=0
(119894 minus 120583119894) (119895 minus 120583119895) 119862 (119894 119895)120590119894120590119895 (20)
119871119900119888119886119897 ℎ119900119898119900119892119890119899119890119894119905119910 = 119871minus1sum119894=0
119871minus1sum119895=0
11 + (119894 minus 119895)2119862 (119894 119895) (21)
23 Firefly Optimization Algorithm The firefly algorithm isa swarm based stochastic search technique [37] The fireflyoptimization algorithm consists of a set of members calledfireflies each firefly represents a candidate solution Themost attractive firefly is considered to be the leader fireflythat leads the other candidates to the best region Theattractiveness is calculated based on the light intensity whichis usually determined by the objective fitness function Theattractiveness between two fireflies 119883119894 and 119883119895 is determinedas follows
120573 (119903119894119895) = 1205730119890minus1205741199032119894119895 (22)
119903119894119895 = radic 119863sum119889=1
(119909119894119889 minus 119909119895119889)2 (23)
where 119863 denotes the problem dimension such that 119863 =1 2 119889 119903119894119895 denotes the distance between 119883119894 and 119883119895Parameter 1205730 denotes the initial attractiveness at 119903 = 0 and120574 denotes the light absorption factor such that 120574 isin [0 1]Each firefly 119883119894 is compared with the other fireflies 119883119895 where119895 isin 1 2 119873 such that 119894 = 119895 and119873 denotes the count of thefireflies If firefly 119883119894 is better (brighter) than 119883119895 then firefly119883119895 moves towards 119883119894 with a step movement formulated asfollows
119883119894119889 (119905 + 1) = 119883119894119889 (119905) + 1205730119890minus1205741199032119894119895119894119895 (119883119895119889 (119905) minus 119883119894119889 (119905))+ 120572120598119894 (24)
where 120598119894 represents uniform a randomly distributed variablesuch that 120598119894 isin [minus05 05] and 120572 denotes the movement stepsuch that 120572 isin [0 1]
3 The Combinational Hybrid System ofEpilepsy Detection from EEG Signal
In this research a hybrid system was proposed to detectboth seizures and seizure-free conditions from a raw EEGsignal Although some investigations focused on the featureextraction level the proposed system was established basedon four main levels This system combined the data fusionapproach with firefly optimization and random forest The119879119876119882119879 was applied for EEG signal decomposition then thefeatures were constructed using a data fusion technique Dueto the large number of features obtained for each subband(119891119890119886119905119906119903119890119904119862119900119906119899119905 times 119869 times 119876) a feature reduction was applied toreduce the features and to obtain a compact set of featuresinstead of the original one The obtained compact set offeatures was fed to a random forest algorithm to obtainthe classification rules and hence used for training Aftertraining the classifier should be able to classify and estimatethe preictal phase The proposed system was divided intothe following four levels of processing and then described indetail as shown in Figure 1
(i) EEG decomposition using TQWT(ii) Feature extraction using data fusion based on single-
channel EEG signal and co-occurrence matrix(iii) Feature reduction using firefly optimization algo-
rithm(iv) Training of random forest classifier to detect the
seizures and seizure-free EEGs
31 EEG Decomposition Using TQWT The preprocessinglevel applies the 119879119876119882119879 decomposition to the input EEGsignal The 119879119876119882119879 converted the continuous EEG signalto discrete potions of data that could be handled moreeffectively This wavelet transformation is used because of itseffectiveness in signal decomposition and its tunability Theobtained subbands using the 119879119876119882119879 provided a significantdifference between the seizure-free and the epileptic seizureof the EEG signals as shown in Figure 2 The subfiguresdenoted by (a) (b) (c) and (d) visualize the histogram of thefirst second third and fourth subbands of the seizure-freeclass The remaining subfigures denoted by (e) (f) (g) and(h) represent the histogram of the epileptic seizure class forthe same subbands The values of the extracted subbands ofthe second class are about ten times stronger than the firstclass that prove the efficiency of this decomposition
32 Feature Extraction Using Data Fusion Based on a Single-Channel EEG and a Co-Occurrence Matrix In the firstperspective the EEG signal was described as a nonstationarytime series After the decomposition of the EEG signal usingthe TQWT a feature extraction process was performed toobtain significant characteristics fromeachTQWTsubbandsThe extracted features were categorized into three maingroups The first group determines the statistical character-istics in the time domain The mean standard deviationvariance Shannon entropy and approximate entropy werecalculated in the first group as formulated from (5) to (9)
International Journal of Biomedical Imaging 5
Image Extract
chaotic features Apply TQWT for the signal decomposition Train the random forest
classifier
EEG Signal
Extractstatistical features Feature reduction using
firefly algorithm End Constrcut feature space using data
fusion
Classifiy and estimate seizures using the trained
model
Single EEG Channel
Extract power spectrum
features
Extract Co-occurrence
matrix
Input Level 1 Level 2 Level 3 Level 4
Figure 1 Block diagram of the combinational hybrid system of epilepsy detection from EEG signal with 4 levels of processing
2500
2000
1500
1000
500
0
occu
ranc
e
minus20 0 20
V
(a)
0 20 40minus20minus40
V
0
500
1000
1500
2000
2500
3000
occu
ranc
e
(b)
0
200
400
600
800
1000
1200
1400
0 50 100minus50
V
occu
ranc
e
(c)
0
200
400
600
800
1000
1200
0 200 400minus200
V
occu
ranc
e
(d)
0
500
1000
1500
2000
2500
0 200minus200
V
occu
ranc
e
(e)
2500
2000
1500
1000
500
00 200 400minus200minus400
V
occu
ranc
e
(f)
0 500minus500
V
0
200
400
600
800
1000
1200
occu
ranc
e
(g)
0
200
400
600
800
1000
1200
0 500 1000minus500minus1000
V
occu
ranc
e
(h)
Figure 2 Illustration of seizure-free and epileptic seizures of subbands obtained from TQWT with Q = 1 r = 3 J = 3 Figures (a) (b) (c)and (d) represent the first second third and fourth subbands obtained from seizure-free signals Figures (e) (f) (g) and (h) represent thefirst second third and fourth subbands obtained from the epileptic seizures
This group indicates some statistical information obtainedfrom the time domain of TQWT subband The secondgroup of features determines a power spectrum analysisof obtained subbands The discrete Fourier transformations(DFT) were applied to convert these subbands into a fre-quency domain Then the power spectrum features wereextractedThe second group consists of the spectral centroidspectral speed spectral flatness spectral slope and spectralentropy as formulated from (10) to (14) The power spectrumanalysis represents an effectivemethod to study the frequencybehavior of the signal The last group of features performs achaotic analysis of each subband Because of the nonlinearityof the EEG signals a nonlinear analysis is required One ofthe best analyses used for this issue is the chaotic analysisIn this analysis the Higuchi fractal dimension (HFD) Hurst
exponent and Katz fractal exponent were computed for eachsubband as formulated from (15) to (18)
In the second perspective the input EEG signal wasconverted to a gray image Then a co-occurrence matrix wascomputed to obtain the gray levels of the image Then thetextures of contrast correlation energy and homogeneitywere calculated from this matrix to represent the statisticalmeasures of the image as formulated from (19) to (21) Finallyafter computing the feature space a data fusion was appliedto merge all of these features and create a single dataset
33 Feature Reduction Using Firefly Optimization AlgorithmIn this section a feature reduction algorithm based onthe firefly algorithm is proposed [37 38] This algorithmimplements a chaotic movement simulated annealing (SA)
6 International Journal of Biomedical Imaging
(1) InputThe features matrix 119865119898(2) OutputThe optimal solution 119892119887119890119904119905(3) Initialize the firefly swam using a logistic chaotic map(4) Evaluate each firefly 119892119894 using the fitness function f (x)(5) Select the best firefly 119892119887119890119904119905 and the worst one 119892119908119900119903119904119905(6) while termination condition are not reached do(7) Declare an alternative leader firefly as 119878119887119890119904119905 with a competitive fitness and located in different region(8) Obtain an offspring solution 119892119887119890119904119905 using 119878119860(9) for all (firefly 119894 and 119891119894 = 119892119908119900119904119903119905) do(10) for all (firefly 119895 and 119891119895 = 119892119908119900119904119903119905) do(11) if (119868119895 gt 119868119894) then(12) Enhance firefly 119895 using Equation (25) to obtain better offspring candidate firefly denoted 1199091015840119895(13) Replace firefly 119895 with the offspring 1199091015840119895(14) Move the firefly 119895 towards the neighboring and global optimal solutions using Equation (26)(15) end if(16) end for(17) end for(18) Update the worst solution 119892119908119900119903119904119905(19) if 119891(1198921015840119887119890119904119905) gt 119891(119892119887119890119904119905) then(20) 119892119887119890119904119905 larr997888 1198921015840119887119890119904119905(21) end if(22) Rank all fireflies and update the best 119892119887119890119904119905 and worst 119892119908119900119903119904119905 solutions(23) end while(24) return 119892119887119890119904119905Algorithm 1 The pseudocode of the proposed feature reduction algorithm based on firefly optimization and SA
to produce efficient offspring candidates andmemory aware-ness of the best and worst solutions to improve the searchdiversity and prevent local optima The firefly population israndomly initialized using a chaotic logistic map to ensurethe diversely of the candidates and the randomness of eachfirefly Afterwards the fitness function is computed for eachcandidate and identifies both the best and worst solutions as119892119887119890119904119905 and 119892119908119900119904119903119905 respectively For each iteration an alternativecandidate is declared as 119878119887119890119904119905 with a competitive fitnessand located in a different region Both of the best and thealternative candidate sets are used to lead weak solutions toreach the optimal region and prevent local optima problemThe mean of the leader firefly and the alternative one isenhanced using SA algorithm to obtain a better solution1198921015840119887119890119904119905 The improved local and global solutions are used toguide the low lightness fireflies to move towards that strongerlightness The algorithm is repeated until the maximumnumber of iterations is reached or a termination criterion isachieved The behavior of the proposed algorithm is deter-mined by some properties namely the objective function theattractiveness movement step and population diversity Theproposed algorithm is demonstrated in Algorithm 1
TheObjective FunctionThis function is used to evaluate eachcandidate in the algorithm and defined as follows
119891 (119909) = 1199081 times 119886119888119888119906119903119886119888119910 (119909) + 1199082number of features (25)
where 1199081 and 1199082 represent the weights of the classifi-cation accuracy and the number of the selected features
respectively The values of 1199081 and 1199082 are set to 09 and 01respectively as a recommended by [38]
The Attractiveness Movement Step The proposed algorithmused a chaotic logistic map to initialize the firefly populationand hence increases the diversity and avoid local optimaAfter obtaining the global best solution 119892119887119890119904119905 an alternativeleader firefly 119878119887119890119904119905 is declared with a competitive fitness butlocated in a different region Since both leaders are morelikely to discover distinctive search regions this strategyreduces the probability of being trapped in the local optimaIn addition the optimal offspring of themean positions of thetwo leaders and the neighboring brighter candidates are usedto lead the search process and guide the solutions with lowerlight intensity to move towards the optimal region
119909119894 = 119909119894 + 1205730119862119896 (1199091015840119895 minus 119909119894) + 119862119896120576 (1198921015840119887119890119904119905 minus 119909119894) + 1205721015840times sign [119903119886119899119889 minus 05] (26)
1199091015840119895 = 119909119895 + 1205901 (27)
1198921015840119887119890119904119905 = 119898119890119886119899 (119892119887119890119904119905 + 119878119887119890119904119905) + 1205902 (28)
where 1199091015840119895 denotes the offspring candidate with a brighterneighboring solution119909119895 is defined by the SA as shown in (26)and 1198921015840119887119890119904119905 represents the fitter offspring solutions of the meanof the leader firefly and the alternative one as formulated in(27) It worth mentioning that the values of 1205901 and 1205902 aretwo random variables set using the Gaussian distributionThe movement step of the firefly is determined as shown in
International Journal of Biomedical Imaging 7
(1) Input 119894119905119890119903119886119905119894119900119899119904119898119886119909 119892119898119890119886119899 119879119898119886119909(2) OutputThe optimal offspring 119878119887119890119904119905(3) 119878119888 = 119862119903119890119886119905119890119868119899119894119905119878119900119906119897119905119894119900119899119904(119892119898119890119886119899)(4) 119878119887119890119904119905 = 119878119888(5) while (119894 le 119894119905119890119903119886119905119894119900119899119904 119898119886119909) do(6) 119878119894 = 119862119903119890119886119905119890119873119890119894119892ℎ119887119900119903119878119900119897119906119905119894119900119899(119878119888)(7) 119879119888 = 119862119886119897119888119906119897119886119905119890119879119890119898119901119890119903119886119905119906119903119890(119894 119879119898119886119909)(8) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119888)) then(9) 119878119888 = 119878119894(10) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119887119890119904119905)) then(11) 119878119887119890119904119905 = 119878119894(12) end if(13) else if (exp((119888119900119904119905(119878119887119890119904119905) minus 119888119900119904119905(119878119894))119879119888) ge 120590) then(14) 119878119888 = 119878119894(15) end if(16) 119894 = 119894 + 1(17) end while(18) return 119878119887119890119904119905Algorithm 2 The pseudocode of the generation of offspring solutions using SA
(28) where 119862119870 represents the chaotic map variable in themovement step and 120576 denotes the randomized vector definedin the traditional firefly algorithm Parameter 1205721015840 denotes anadaptive step initialized to 05 to control the diversity of thesearch process
Offspring Generation Using SAThe proposed algorithm usedSA for generating better candidates to enhance the searchprocess as much as possible as shown in Algorithm 2 TheSA accepts both of the best solution 119892119887119890119904119905 and the alternativesolution 119878119887119890119904119905 as main inputs then the traditional SA isapplied The better solution generated is accepted by defaultaccording to the SA heuristics On the other hand theweaker solution should be accepted with specific probabilityas shown in (29) where Δ119891 denotes the difference of thefitness (energy) between to candidates and 119879119888 denotes thecurrent temperature A simple linear cooling mechanism isused to control the value of the temperature
119901119909119895 = exp(Δ119891119879119888 ) (29)
Population Diversity In each iteration the worst solution isdetected after the ranking process as shown in Algorithm 1The remaining solutions are guided by the average positionobtained from (30) where 120590 denotes a random value obtainedby Gaussian map
119909119908119900119903119904119905119895 = 119892119887119890119904119905 + 1198781198871198901199041199052 + 120590 (119909119908119900119903119904119905119895 119892119887119890119904119905 + 1198781198871198901199041199052 ) (30)
34 Classification and Learning The random forest (RF) isa successful ensemble approach used in supervised machinelearning to solve classification or regression problems [39]It consists of a collection of decision trees that could actas a single classifier with multiple classification methodsor a method that has several variables Several subsets ofthe training data are supplied to each tree to achieve the
most stable tree classification that results in a generalizedexperience of the classifier The original dataset is dividedinto two parts The first part is used to train each tree bybootstrapping technique The other part is used to evaluatethe accuracy of the classification Each tree is allowed to reachthe maximum depth without tree pruning to obtain a highvariance classifier The splitting process remains until onlyone instance of a single class is dropped from any leaf nodeor a predefined termination condition is achieved Whenthe forest is established the number of subsets remained asa constant The obtained route of traversal from the rootnode to the leaf node is applied to the new instances orthe unlabeled instance for classification The final decisionfor classifying a new instance is provided by determiningeach class that has the most votes from every decision treeThe random forest performs slightly better when comparedwith other classifiers such as discriminant analysis SVM andartificial neural network (ANN) [15 39]
35 Performance Evaluation Various performance formulaswere used to evaluate the effectiveness of a classifierThe sen-sitivity (SEN) or recall specificity (SPEC) accuracy (ACC)F-measure Matthewrsquos correlation coefficient (MCC) andreceiver operating characteristics (ROC) are used to evaluatethe efficiency of the random forest classifier [40ndash42] Theseparameters are defined as follows
119878119864119873 = 119879119875119879119875 + 119865119873 times 100 (31)
119878119875119864119862 = 119879119873119879119873 + 119865119875 times 100 (32)
119860119862119862 = 119879119875 + 119879119873119879119875 + 119879119873 + 119865119875 + 119865119873 times 100 (33)
119865 minus 119898119890119886119904119906119903119890 = 21198791198752119879119875 + 119865119875 + 119865119873 times 100 (34)
8 International Journal of Biomedical Imaging
119872119862119862= (119879119875 times 119879119873) minus (119865119875 times 119865119873)radic(119879119875 + 119865119873) (119879119875 + 119865119875) (119879119873 + 119865119873) (119879119873 + 119865119875)times 100
(35)
119879119875 and 119879119873 represent the total number of an epileptic seizureand seizure-free signals classified correctly respectively Sim-ilarly 119865119875 and 119865119873 represent the total number of epilepticseizures and seizure-free signals classified incorrectly respec-tively Cross-validation has also been used to ensure theclassifier reliability and effectiveness The original dataset issplit into 119896 folds (subsets) for both training and testing Inthis strategy 119896 minus 1 folds were selected randomly to train theclassifier and the remaining folds are used to testing Theoverall performance is calculated as the average of each foldIn this work the experiment is repeated 10 times with tenfoldcross-validation
4 Results and Discussion
The benchmark dataset used in this investigation wasacquired by the University of Bonn [28]The dataset containsthree different categories ie preictal healthy and ictalrecorded using a single channel for a 236 s duration Bothnormal and preictal conditions were collected from 200 casestudies and 100 for the ictal state The normal condition isacquired from five healthy volunteers using the international10ndash20 system standard with each volunteer in a relaxed-awake state with eyes open and closed (100 cases per each set)[28] The ictal data were collected from five patients duringtheir epileptic seizures The preictal represents the EEG datacollected from the same five patients with no seizures It isworth mentioning that all EEG signals were acquired using a128 channel amplifier with sampling rate equal 17361Hz [28]Finally a bandpass filter with 05340Hz sim12 dBoctave wasapplied as a filter
The proposed approach for automatic detection of epilep-tic seizures and seizure-free patients was implemented usingMATLAB software A TQWT comparison between theseizure-free and the epileptic seizures patients are shown inFigure 3 with 119876 = 1 119903 = 3 and 119869 = 3 It can be observedthat both of the amplitudes and the frequency of the epilepticseizure are much higher than the healthy one Moreover theoscillatory behavior of the epileptic patient is higher than thehealthy one The value of the parameter 119903 is set to three toprevent any excessive ringing of the wavelet as suggested by[22] MATLAB for the TQWT toolbox is available for publicaccess at httpeewebpolyeduiselesniTQWT
After the construction of the dataset the feature reduc-tion was applied using the firefly algorithm to remove theredundant and irrelevant features The number of the datasegments was determined by the parameters 119876 119903 and 119869 Bytuning these parameters the number of the data segmentswas varied and thus the training process was adopted Thetrial and error approach was used to set the value of theseparameters The experiment was implemented on variousvalues of J to obtain the best level of decomposition The
performance measures were calculated for each level ofdecomposition As shown in Figure 4 the best value of thevariable 119869 was from two to three Moreover the best value ofthe parameter119876was found to be one as shown in Figure 5 Allthe variables remained constant while changing the Q-factorto obtain the best value
Then the firefly algorithm with a population size of 20fireflies mutation probability of 001 and light absorptionequal to 01 was applied to reduce the feature setThe result ofthe compact set of features is obtained after 100 iterations andshown in Figure 6 are the number of features obtained by thefirefly algorithm and their corresponding accuracy precisionspecificity and the recall In the last step the feature set isfed to a random forest classifier to obtain the seizure-free andseizure conditionsThe random forest classifier obtained 98accuracy 97 precision 97 specificity 98 recall 98 F-measure and 95 MCC at the third level of decompositionand the Q-factor equals 1
The firefly algorithm reduced the search space into threefeatures These features could replace the original datasetwhich minimizes the processing time The compact set offeature contains the 119878119879119863 119860119901119864119899 and the 119870119860119879119885 the classi-fication rules are based on these features The classificationrules obtained by the proposed system are shown in Figure 7where the decision tree consists of 4 leaves and 7 nodes andthe final decision represents the seizure-free and the epilepticseizure denoted by 01 respectively
A comparative study of the proposed hybrid epilepsydetection approach and other existing classification systemshas been performed in terms of the total accuracy A novelmethod based on the EMDs was proposed to detect theepileptic seizures of epilepsy This method used the Hilberttransformation of IMFs obtained by EMD process thatprovided an analytic signal representation of IMFs [12] Theclassification rules obtained by this method achieved anaccuracy of 90 The usage of frequency domain featuresand Burgrsquos method obtained 9311 accuracy with SVMclassifier [43] Nonlinear features have been used with aGaussianmixturemodel classifier and achieved 95accuracy[44] A decision tree classifier is used with energy fractaldimension and sample entropy and provided 957 [45]A combination of ApEn and the Hurst exponent has beenused to detect the diagnostics of epilepsy and produced 965accuracy with SVM classifier and ANN [46] In [47] aneigensystem based method was proposed cooperated withMultiple Layer Perceptron to classify the epileptic seizureshealthy and the seizure-free This approach provided anaverage accuracy of 975 The EMD methods for epilepsydetection achieved 9775 accuracy [6]The Kraskov entropyis also combinedwith the SVMclassifier and provided 9775[22] An automated diagnostics system based on a set ofentropies and fuzzy Sugeno classifier (FSC) achieved accu-racy up to 98 [19] The work presented by [48] developeda method for the epilepsy detection using the EMD Thegenerated IMFs using the EMD were represented as a set ofamplitude and frequency modulated (AMndashFM) signals Thetwo bandwidths namely amplitude modulation bandwidthand frequency modulation bandwidth calculated from theanalytic IMFs have been fed to LS-SVM for classifying
International Journal of Biomedical Imaging 9
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(a) EEG Sample for patient with seizure-free
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(b) EEG Sample for patient with seizures
Figure 3 TQWT decomposition of seizure-free and seizure EEG signals with 119876 = 1 119903 = 3 119895 = 3
1 2 3 4 5 10LEVELS OF DECMPOSISTION (J)
ACCPrecision
SPECRecall
000
020
040
060
080
100
PERC
ENTA
GE
Figure 4 Performance measures of the proposed approach withvaried levels of decomposition
08082084086088
09092094096098
1
1 2 3 5 7 9 10
PERC
ENTA
GE
Q
ACCPrecision
SPECRecall
Figure 5 Performance measures of the proposed approach withvaried levels of decomposition
seizure and nonseizure EEG signals This method achieved9818 average accuracy The LBP-based methods have beencombined with Gabor filter for texture extraction from theEEG signal A k-nearest neighbor classifier was applied andobtained an accuracy of 983 [14] The proposed methodconfirmed its superiority in the total accuracy compared tothe other systems The results which prove the superiority ofthe proposed method compared to the other existed systemsis demonstrated in Figure 8
ACCPrecision
SPECRecall
09091092093094095096097098099
1
1 2 3 4 5PE
RCEN
TAG
ENUMBER OF FEATURES
Figure 6 Performance measures of the original and compactfeatures set
STD
lt= 2873 gt 2873
0
0
ApEn
lt= 016 gt 016
Katz 1
1
lt= 146 gt 146
Figure 7 The classification rules obtained from the proposedsystem to classify the seizure-free (0) and the epileptic seizure (1)signals
Once the system detects the preictal phase the clinicreceives a notification about that patientThemain advantageof the hybrid system from the clinical point of view could besummarized as follows
(i) Classification and detecting of the epileptic seizuresand seizure-free signals from the EEG signal automat-ically
10 International Journal of Biomedical Imaging
90
931195 957 965
975 9775 9775 98 9813 983 99
Baja
j et a
l 20
13
Faus
t et a
l 20
10
Acha
rya e
t al
2009
Mar
tis et
al 2
013
Yuan
et al
201
1
Nag
hsh-
Nilc
hi et
al 2
010
Pach
ori e
t al
2014
Patid
ar et
al 2
017
Acha
rya e
t al
2012
Baja
j et a
l 20
12
Kum
ar et
al 2
015
e p
ropo
sed
met
hod
PROPOSED SYSTESMS USED TO FOR THE COMPARSION
8486889092949698
100
ACCU
RACY
IN P
ERCE
NTA
GE
()
Figure 8 The total accuracy () of various methods employed to detect the disorder of the epilepsy
(ii) The detection of the preictal phase from studying thehealthy and ictal phase of each patient
(iii) The detection of the preictal phase that provided theability to send warning alert to the physicians toprepare the medical assessment for the patient
(iv) The proposed method being robust and reliable as itsperformance was benchmarked using 10-fold cross-validation
(v) The dynamic behavior obtained from the fireflyalgorithm which made the system adaptive to manyfeatures according to each case study
(vi) A few sets of parameters required to analyze the EEGsignal typical three variables (119876 119903 119869)
The limitations of this research were summarized as follows
(i) The limited number of the studied subjects (typically100 per class)
(ii) The diagnosis process that may be reduced because ofdepending on some additional software prerequisites
5 Conclusion
In this paper an automated intelligent CAD tool has beenproposed to classify and detect epileptic seizures and seizure-free EEG signals This method provided an EEG signalanalysis using a hybrid data fusion method The data fusionmethod combined the collected features from two differentperspectives In the first perspective the EEG signal wasconsidered as an image Then the image was converted toa gray image and the GLCM was obtained to extract thetextures of the image such as contrast correlation power andhomogeneity In the second perspective the EEG signal wasdivided into smaller segments using the TQWT to extractthe time and frequency features A set of statistical nonlinear(chaotic) and power spectrum features were obtained from
each segment After the dataset was constructed a featurereduction algorithmbased on firefly optimizationwas used toreduce the irrelevant features and remove redundancy Thenan RFwas trained to classify and predict the epileptic seizuresand seizure-free EEG signals from the dataset The experi-mental results showed that the proposed method achieveda satisfactory degree of 99 accuracy 97 precision 97specificity 98 recall 98 F-measure and 95 MCC at thethird level of decomposition
Data Availability
The dataset used to support the findings of this research wasincluded in the article and can be found as a citation to ldquoRGAndrzejak K Lehnertz F Mormann C Rieke P DavidCE Elger Indications ofNonlinearDeterministic and Finite-Dimensional Structures in Time Series of Brain ElectricalActivity Dependence on Recording Region and Brain StatePhysical Review E 64 (2001) 061907rdquo
Conflicts of Interest
The authors declare that they have no conflicts of interest
References
[1] Epilepsy Fact Sheet 2018 httpwwwwhointmediacentrefactsheetsfs999en
[2] U R Acharya S V Sree G Swapna R J Martis and J S SurildquoAutomated EEG analysis of epilepsy a reviewrdquo Knowledge-Based Systems vol 45 pp 147ndash165 2013
[3] A Puce and M S Hamalainen ldquoA review of issues relatedto data acquisition and analysis in EEGMEG studiesrdquo BrainSciences vol 7 no 6 2017
[4] V Joshi R B Pachori and A Vijesh ldquoClassification of ictaland seizure-free EEG signals using fractional linear predictionrdquoBiomedical Signal Processing and Control vol 9 pp 1ndash5 2014
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
International Journal of Biomedical Imaging 3
validation of the proposed method A numerical experimenthas been implemented and a comparative study presenteda promising efficiency of the proposed system regarding theoverall accuracy sensitivity and specificity
The remainder of this manuscript is organized as followsthe preliminaries concepts were introduced in Section 2Section 3 introduced the combinational hybrid system ofthe epilepsy detection The experiment and discussion werepresented in Section 4 Lastly the paper was concluded inSection 5
2 Preliminary Knowledge
21 Tunable Q-Wavelet Transformation (TQWT) The tun-ability of the Q-factor provided a proficient method to adoptthe wavelet transformation [25] The TQWT have threeinputsQ-factor denoted by119876 which determines the numberof oscillations of the wavelet the number of the oversamplingrate which is denoted by 119903 andwhich determines the numberof the overlapping frequency responses and the number ofstages of decomposition denoted by 119869 For each decomposi-tion stage the target signal 119904[119899]with a sample rate of 119891119904 couldbe represented by low-pass and high-pass subbands withsampling frequencies of120572119891119904 and120573119891119904 respectively where120572 and120573 are the parameters of signal scaling The low-pass subbandis presented by low-pass filter 1198670(120596) and low-pass scaling119871119875119878(120572) Similarly the high-pass subband 1205961 is produced by1198671(120596) and 119867119875119878(120573) The low-pass and high-pass subbandsignals are formulated as follows
H1198880 (120596)=
0 120572120587 le |120596| le 120587120579(120596 + (120573 minus 1) 120587120572 + 120573 minus 1 ) (1 minus 120573) 120587 le |120596| lt 120572120587
1 |120596| lt (1 minus 120573) 120587
(1)
H1 (120596) =
0 |120596| lt (1 minus 120573) 120587120579( 120572120587 minus 120596120572 + 120573 minus 1) (1 minus 120573) 120587 le |120596| lt 1205721205871 120572120587 le |120596| le 120587
(2)
where 120579(120596) could be defined as follows
120579 (120596) = 05 (1 + cos (120596)) (2 minus cos (120596))05 |120596| le 120587 (3)
Both of 119903 and 119876 could be represented as filter-bank variables120572 and 120573 as follows
119903 = 1205731 minus 120572 119876 = 2 minus 120573120573 (4)
22 Feature Sets The feature sets used in this research aregrouped into four main groups which are statistical powerspectrum chaotic features and gray-level co-occurrencematrix (GLCM)The first group contains a set of five features
calculated from the time domain of the input signal Thisfeature set containsmean (120583) standard deviation (119878119879119863) vari-ance (var) Shannon entropy (119867) and approximate entropy(ApEn) The mathematical formulation of each feature isshown as follows [28ndash31]
120583 (119909) = 1119873119873sum119894=1
119909119894 (5)
119878119879119863 (119909) = radic 1119873 minus 1119873sum119894=1
(119909119894 minus 120583119909)2 (6)
var (119909) = 1119873 minus 1119873sum119894=1
1003816100381610038161003816119909119894 minus 12058310038161003816100381610038162 (7)
119867(119883) = 119873sum119894=1
119875 (119909119894) log (119875 (119909119894)) (8)
119860119901119864119899 (x) = 120601119898 (119903) minus 120601119898+1 (119903) (9)
The second set of features calculates the power spectrum ofthe input signal based on the frequency domain analysisThis feature set contains spectral centroid (119878119862) spectral speed(119878119878) spectral flatness (119878119865) spectral slope (119878119878119868) and spectralentropy (119875119878119864) where 119884(119902) denotes the for the discreteFourier transformation of the input signal 119891(119899) The math-ematical formulation of each feature is shown as follows [32]
119878119862 = sum119872minus1119902=0 119902 1003816100381610038161003816119884 (119902)1003816100381610038161003816sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816 (10)
119878119878 = sum119872minus1119902=0 (119902 minus 119878119862)2 1003816100381610038161003816119884 (119902)1003816100381610038161003816sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816 (11)
119878119865 = prod119872minus1119902=0 1003816100381610038161003816119884 (119902)10038161003816100381610038161119872(1119872)sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816 (12)
SSI = 119872sum119872minus1119902=0 119891119898 1003816100381610038161003816119884 (119902)1003816100381610038161003816 minus sum119872minus1119902=0 119891119898 sdot sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816119872sum119872minus1119902=0 1198912119898 minus (sum119872minus1119902=0 1003816100381610038161003816119884 (119902)1003816100381610038161003816)2 (13)
119875119878119864 = minus 119899sum119894=1
119884 (119902)sum119899119894=1 119884 (119902) ln(119884 (119902)sum119899119894=1 119884 (119902)) (14)
The third set of features contains chaotic measures to obtainthe dynamic behavior of the EEG signal This set includesHiguchirsquos fractal dimension (119867119865119863) Hurst exponent (119867119903)and Katz fractal exponent (119870119860119879119885) These features areformulated as follows [33ndash36]
119867119865119863 = ln( 119896119871 (119896)) (15)
119864[119877 (119899)119878 (119899) ] = 119888119899119867119903 (16)
119870119860119879119885 = log (119899)log (119899) + log (119889119871) (17)
4 International Journal of Biomedical Imaging
The final set of features consists of statistical measuresof an image represented as matrices called gray-level co-occurrence matrix (GLCM) where 119862(119894 119895) represents an entryin co-occurrence matrix and 119894 119895 = 0 1 2 119871 minus 1 where119871 is the number of gray levels in the image Those matricesrepresent the spatial dependencies between the gray levels ofimage reflecting the structure of the underlying texture Afterthe normalization of these matrices the contrast correlationenergy and homogeneity are computed as follows
119864119899119890119903119892119910 = 119871minus1sum119894=0
119871minus1sum119895=0
119862 (119894 119895)2 (18)
119862119900119899119905119903119886119904119905 = 119871minus1sum119894=0
119871minus1sum119895=0
(119894 minus 119895)2 119862 (119894 119895) (19)
119862119900119903119903119890119897119886119905119894119900119899 = 119871minus1sum119894=0
119871minus1sum119895=0
(119894 minus 120583119894) (119895 minus 120583119895) 119862 (119894 119895)120590119894120590119895 (20)
119871119900119888119886119897 ℎ119900119898119900119892119890119899119890119894119905119910 = 119871minus1sum119894=0
119871minus1sum119895=0
11 + (119894 minus 119895)2119862 (119894 119895) (21)
23 Firefly Optimization Algorithm The firefly algorithm isa swarm based stochastic search technique [37] The fireflyoptimization algorithm consists of a set of members calledfireflies each firefly represents a candidate solution Themost attractive firefly is considered to be the leader fireflythat leads the other candidates to the best region Theattractiveness is calculated based on the light intensity whichis usually determined by the objective fitness function Theattractiveness between two fireflies 119883119894 and 119883119895 is determinedas follows
120573 (119903119894119895) = 1205730119890minus1205741199032119894119895 (22)
119903119894119895 = radic 119863sum119889=1
(119909119894119889 minus 119909119895119889)2 (23)
where 119863 denotes the problem dimension such that 119863 =1 2 119889 119903119894119895 denotes the distance between 119883119894 and 119883119895Parameter 1205730 denotes the initial attractiveness at 119903 = 0 and120574 denotes the light absorption factor such that 120574 isin [0 1]Each firefly 119883119894 is compared with the other fireflies 119883119895 where119895 isin 1 2 119873 such that 119894 = 119895 and119873 denotes the count of thefireflies If firefly 119883119894 is better (brighter) than 119883119895 then firefly119883119895 moves towards 119883119894 with a step movement formulated asfollows
119883119894119889 (119905 + 1) = 119883119894119889 (119905) + 1205730119890minus1205741199032119894119895119894119895 (119883119895119889 (119905) minus 119883119894119889 (119905))+ 120572120598119894 (24)
where 120598119894 represents uniform a randomly distributed variablesuch that 120598119894 isin [minus05 05] and 120572 denotes the movement stepsuch that 120572 isin [0 1]
3 The Combinational Hybrid System ofEpilepsy Detection from EEG Signal
In this research a hybrid system was proposed to detectboth seizures and seizure-free conditions from a raw EEGsignal Although some investigations focused on the featureextraction level the proposed system was established basedon four main levels This system combined the data fusionapproach with firefly optimization and random forest The119879119876119882119879 was applied for EEG signal decomposition then thefeatures were constructed using a data fusion technique Dueto the large number of features obtained for each subband(119891119890119886119905119906119903119890119904119862119900119906119899119905 times 119869 times 119876) a feature reduction was applied toreduce the features and to obtain a compact set of featuresinstead of the original one The obtained compact set offeatures was fed to a random forest algorithm to obtainthe classification rules and hence used for training Aftertraining the classifier should be able to classify and estimatethe preictal phase The proposed system was divided intothe following four levels of processing and then described indetail as shown in Figure 1
(i) EEG decomposition using TQWT(ii) Feature extraction using data fusion based on single-
channel EEG signal and co-occurrence matrix(iii) Feature reduction using firefly optimization algo-
rithm(iv) Training of random forest classifier to detect the
seizures and seizure-free EEGs
31 EEG Decomposition Using TQWT The preprocessinglevel applies the 119879119876119882119879 decomposition to the input EEGsignal The 119879119876119882119879 converted the continuous EEG signalto discrete potions of data that could be handled moreeffectively This wavelet transformation is used because of itseffectiveness in signal decomposition and its tunability Theobtained subbands using the 119879119876119882119879 provided a significantdifference between the seizure-free and the epileptic seizureof the EEG signals as shown in Figure 2 The subfiguresdenoted by (a) (b) (c) and (d) visualize the histogram of thefirst second third and fourth subbands of the seizure-freeclass The remaining subfigures denoted by (e) (f) (g) and(h) represent the histogram of the epileptic seizure class forthe same subbands The values of the extracted subbands ofthe second class are about ten times stronger than the firstclass that prove the efficiency of this decomposition
32 Feature Extraction Using Data Fusion Based on a Single-Channel EEG and a Co-Occurrence Matrix In the firstperspective the EEG signal was described as a nonstationarytime series After the decomposition of the EEG signal usingthe TQWT a feature extraction process was performed toobtain significant characteristics fromeachTQWTsubbandsThe extracted features were categorized into three maingroups The first group determines the statistical character-istics in the time domain The mean standard deviationvariance Shannon entropy and approximate entropy werecalculated in the first group as formulated from (5) to (9)
International Journal of Biomedical Imaging 5
Image Extract
chaotic features Apply TQWT for the signal decomposition Train the random forest
classifier
EEG Signal
Extractstatistical features Feature reduction using
firefly algorithm End Constrcut feature space using data
fusion
Classifiy and estimate seizures using the trained
model
Single EEG Channel
Extract power spectrum
features
Extract Co-occurrence
matrix
Input Level 1 Level 2 Level 3 Level 4
Figure 1 Block diagram of the combinational hybrid system of epilepsy detection from EEG signal with 4 levels of processing
2500
2000
1500
1000
500
0
occu
ranc
e
minus20 0 20
V
(a)
0 20 40minus20minus40
V
0
500
1000
1500
2000
2500
3000
occu
ranc
e
(b)
0
200
400
600
800
1000
1200
1400
0 50 100minus50
V
occu
ranc
e
(c)
0
200
400
600
800
1000
1200
0 200 400minus200
V
occu
ranc
e
(d)
0
500
1000
1500
2000
2500
0 200minus200
V
occu
ranc
e
(e)
2500
2000
1500
1000
500
00 200 400minus200minus400
V
occu
ranc
e
(f)
0 500minus500
V
0
200
400
600
800
1000
1200
occu
ranc
e
(g)
0
200
400
600
800
1000
1200
0 500 1000minus500minus1000
V
occu
ranc
e
(h)
Figure 2 Illustration of seizure-free and epileptic seizures of subbands obtained from TQWT with Q = 1 r = 3 J = 3 Figures (a) (b) (c)and (d) represent the first second third and fourth subbands obtained from seizure-free signals Figures (e) (f) (g) and (h) represent thefirst second third and fourth subbands obtained from the epileptic seizures
This group indicates some statistical information obtainedfrom the time domain of TQWT subband The secondgroup of features determines a power spectrum analysisof obtained subbands The discrete Fourier transformations(DFT) were applied to convert these subbands into a fre-quency domain Then the power spectrum features wereextractedThe second group consists of the spectral centroidspectral speed spectral flatness spectral slope and spectralentropy as formulated from (10) to (14) The power spectrumanalysis represents an effectivemethod to study the frequencybehavior of the signal The last group of features performs achaotic analysis of each subband Because of the nonlinearityof the EEG signals a nonlinear analysis is required One ofthe best analyses used for this issue is the chaotic analysisIn this analysis the Higuchi fractal dimension (HFD) Hurst
exponent and Katz fractal exponent were computed for eachsubband as formulated from (15) to (18)
In the second perspective the input EEG signal wasconverted to a gray image Then a co-occurrence matrix wascomputed to obtain the gray levels of the image Then thetextures of contrast correlation energy and homogeneitywere calculated from this matrix to represent the statisticalmeasures of the image as formulated from (19) to (21) Finallyafter computing the feature space a data fusion was appliedto merge all of these features and create a single dataset
33 Feature Reduction Using Firefly Optimization AlgorithmIn this section a feature reduction algorithm based onthe firefly algorithm is proposed [37 38] This algorithmimplements a chaotic movement simulated annealing (SA)
6 International Journal of Biomedical Imaging
(1) InputThe features matrix 119865119898(2) OutputThe optimal solution 119892119887119890119904119905(3) Initialize the firefly swam using a logistic chaotic map(4) Evaluate each firefly 119892119894 using the fitness function f (x)(5) Select the best firefly 119892119887119890119904119905 and the worst one 119892119908119900119903119904119905(6) while termination condition are not reached do(7) Declare an alternative leader firefly as 119878119887119890119904119905 with a competitive fitness and located in different region(8) Obtain an offspring solution 119892119887119890119904119905 using 119878119860(9) for all (firefly 119894 and 119891119894 = 119892119908119900119904119903119905) do(10) for all (firefly 119895 and 119891119895 = 119892119908119900119904119903119905) do(11) if (119868119895 gt 119868119894) then(12) Enhance firefly 119895 using Equation (25) to obtain better offspring candidate firefly denoted 1199091015840119895(13) Replace firefly 119895 with the offspring 1199091015840119895(14) Move the firefly 119895 towards the neighboring and global optimal solutions using Equation (26)(15) end if(16) end for(17) end for(18) Update the worst solution 119892119908119900119903119904119905(19) if 119891(1198921015840119887119890119904119905) gt 119891(119892119887119890119904119905) then(20) 119892119887119890119904119905 larr997888 1198921015840119887119890119904119905(21) end if(22) Rank all fireflies and update the best 119892119887119890119904119905 and worst 119892119908119900119903119904119905 solutions(23) end while(24) return 119892119887119890119904119905Algorithm 1 The pseudocode of the proposed feature reduction algorithm based on firefly optimization and SA
to produce efficient offspring candidates andmemory aware-ness of the best and worst solutions to improve the searchdiversity and prevent local optima The firefly population israndomly initialized using a chaotic logistic map to ensurethe diversely of the candidates and the randomness of eachfirefly Afterwards the fitness function is computed for eachcandidate and identifies both the best and worst solutions as119892119887119890119904119905 and 119892119908119900119904119903119905 respectively For each iteration an alternativecandidate is declared as 119878119887119890119904119905 with a competitive fitnessand located in a different region Both of the best and thealternative candidate sets are used to lead weak solutions toreach the optimal region and prevent local optima problemThe mean of the leader firefly and the alternative one isenhanced using SA algorithm to obtain a better solution1198921015840119887119890119904119905 The improved local and global solutions are used toguide the low lightness fireflies to move towards that strongerlightness The algorithm is repeated until the maximumnumber of iterations is reached or a termination criterion isachieved The behavior of the proposed algorithm is deter-mined by some properties namely the objective function theattractiveness movement step and population diversity Theproposed algorithm is demonstrated in Algorithm 1
TheObjective FunctionThis function is used to evaluate eachcandidate in the algorithm and defined as follows
119891 (119909) = 1199081 times 119886119888119888119906119903119886119888119910 (119909) + 1199082number of features (25)
where 1199081 and 1199082 represent the weights of the classifi-cation accuracy and the number of the selected features
respectively The values of 1199081 and 1199082 are set to 09 and 01respectively as a recommended by [38]
The Attractiveness Movement Step The proposed algorithmused a chaotic logistic map to initialize the firefly populationand hence increases the diversity and avoid local optimaAfter obtaining the global best solution 119892119887119890119904119905 an alternativeleader firefly 119878119887119890119904119905 is declared with a competitive fitness butlocated in a different region Since both leaders are morelikely to discover distinctive search regions this strategyreduces the probability of being trapped in the local optimaIn addition the optimal offspring of themean positions of thetwo leaders and the neighboring brighter candidates are usedto lead the search process and guide the solutions with lowerlight intensity to move towards the optimal region
119909119894 = 119909119894 + 1205730119862119896 (1199091015840119895 minus 119909119894) + 119862119896120576 (1198921015840119887119890119904119905 minus 119909119894) + 1205721015840times sign [119903119886119899119889 minus 05] (26)
1199091015840119895 = 119909119895 + 1205901 (27)
1198921015840119887119890119904119905 = 119898119890119886119899 (119892119887119890119904119905 + 119878119887119890119904119905) + 1205902 (28)
where 1199091015840119895 denotes the offspring candidate with a brighterneighboring solution119909119895 is defined by the SA as shown in (26)and 1198921015840119887119890119904119905 represents the fitter offspring solutions of the meanof the leader firefly and the alternative one as formulated in(27) It worth mentioning that the values of 1205901 and 1205902 aretwo random variables set using the Gaussian distributionThe movement step of the firefly is determined as shown in
International Journal of Biomedical Imaging 7
(1) Input 119894119905119890119903119886119905119894119900119899119904119898119886119909 119892119898119890119886119899 119879119898119886119909(2) OutputThe optimal offspring 119878119887119890119904119905(3) 119878119888 = 119862119903119890119886119905119890119868119899119894119905119878119900119906119897119905119894119900119899119904(119892119898119890119886119899)(4) 119878119887119890119904119905 = 119878119888(5) while (119894 le 119894119905119890119903119886119905119894119900119899119904 119898119886119909) do(6) 119878119894 = 119862119903119890119886119905119890119873119890119894119892ℎ119887119900119903119878119900119897119906119905119894119900119899(119878119888)(7) 119879119888 = 119862119886119897119888119906119897119886119905119890119879119890119898119901119890119903119886119905119906119903119890(119894 119879119898119886119909)(8) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119888)) then(9) 119878119888 = 119878119894(10) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119887119890119904119905)) then(11) 119878119887119890119904119905 = 119878119894(12) end if(13) else if (exp((119888119900119904119905(119878119887119890119904119905) minus 119888119900119904119905(119878119894))119879119888) ge 120590) then(14) 119878119888 = 119878119894(15) end if(16) 119894 = 119894 + 1(17) end while(18) return 119878119887119890119904119905Algorithm 2 The pseudocode of the generation of offspring solutions using SA
(28) where 119862119870 represents the chaotic map variable in themovement step and 120576 denotes the randomized vector definedin the traditional firefly algorithm Parameter 1205721015840 denotes anadaptive step initialized to 05 to control the diversity of thesearch process
Offspring Generation Using SAThe proposed algorithm usedSA for generating better candidates to enhance the searchprocess as much as possible as shown in Algorithm 2 TheSA accepts both of the best solution 119892119887119890119904119905 and the alternativesolution 119878119887119890119904119905 as main inputs then the traditional SA isapplied The better solution generated is accepted by defaultaccording to the SA heuristics On the other hand theweaker solution should be accepted with specific probabilityas shown in (29) where Δ119891 denotes the difference of thefitness (energy) between to candidates and 119879119888 denotes thecurrent temperature A simple linear cooling mechanism isused to control the value of the temperature
119901119909119895 = exp(Δ119891119879119888 ) (29)
Population Diversity In each iteration the worst solution isdetected after the ranking process as shown in Algorithm 1The remaining solutions are guided by the average positionobtained from (30) where 120590 denotes a random value obtainedby Gaussian map
119909119908119900119903119904119905119895 = 119892119887119890119904119905 + 1198781198871198901199041199052 + 120590 (119909119908119900119903119904119905119895 119892119887119890119904119905 + 1198781198871198901199041199052 ) (30)
34 Classification and Learning The random forest (RF) isa successful ensemble approach used in supervised machinelearning to solve classification or regression problems [39]It consists of a collection of decision trees that could actas a single classifier with multiple classification methodsor a method that has several variables Several subsets ofthe training data are supplied to each tree to achieve the
most stable tree classification that results in a generalizedexperience of the classifier The original dataset is dividedinto two parts The first part is used to train each tree bybootstrapping technique The other part is used to evaluatethe accuracy of the classification Each tree is allowed to reachthe maximum depth without tree pruning to obtain a highvariance classifier The splitting process remains until onlyone instance of a single class is dropped from any leaf nodeor a predefined termination condition is achieved Whenthe forest is established the number of subsets remained asa constant The obtained route of traversal from the rootnode to the leaf node is applied to the new instances orthe unlabeled instance for classification The final decisionfor classifying a new instance is provided by determiningeach class that has the most votes from every decision treeThe random forest performs slightly better when comparedwith other classifiers such as discriminant analysis SVM andartificial neural network (ANN) [15 39]
35 Performance Evaluation Various performance formulaswere used to evaluate the effectiveness of a classifierThe sen-sitivity (SEN) or recall specificity (SPEC) accuracy (ACC)F-measure Matthewrsquos correlation coefficient (MCC) andreceiver operating characteristics (ROC) are used to evaluatethe efficiency of the random forest classifier [40ndash42] Theseparameters are defined as follows
119878119864119873 = 119879119875119879119875 + 119865119873 times 100 (31)
119878119875119864119862 = 119879119873119879119873 + 119865119875 times 100 (32)
119860119862119862 = 119879119875 + 119879119873119879119875 + 119879119873 + 119865119875 + 119865119873 times 100 (33)
119865 minus 119898119890119886119904119906119903119890 = 21198791198752119879119875 + 119865119875 + 119865119873 times 100 (34)
8 International Journal of Biomedical Imaging
119872119862119862= (119879119875 times 119879119873) minus (119865119875 times 119865119873)radic(119879119875 + 119865119873) (119879119875 + 119865119875) (119879119873 + 119865119873) (119879119873 + 119865119875)times 100
(35)
119879119875 and 119879119873 represent the total number of an epileptic seizureand seizure-free signals classified correctly respectively Sim-ilarly 119865119875 and 119865119873 represent the total number of epilepticseizures and seizure-free signals classified incorrectly respec-tively Cross-validation has also been used to ensure theclassifier reliability and effectiveness The original dataset issplit into 119896 folds (subsets) for both training and testing Inthis strategy 119896 minus 1 folds were selected randomly to train theclassifier and the remaining folds are used to testing Theoverall performance is calculated as the average of each foldIn this work the experiment is repeated 10 times with tenfoldcross-validation
4 Results and Discussion
The benchmark dataset used in this investigation wasacquired by the University of Bonn [28]The dataset containsthree different categories ie preictal healthy and ictalrecorded using a single channel for a 236 s duration Bothnormal and preictal conditions were collected from 200 casestudies and 100 for the ictal state The normal condition isacquired from five healthy volunteers using the international10ndash20 system standard with each volunteer in a relaxed-awake state with eyes open and closed (100 cases per each set)[28] The ictal data were collected from five patients duringtheir epileptic seizures The preictal represents the EEG datacollected from the same five patients with no seizures It isworth mentioning that all EEG signals were acquired using a128 channel amplifier with sampling rate equal 17361Hz [28]Finally a bandpass filter with 05340Hz sim12 dBoctave wasapplied as a filter
The proposed approach for automatic detection of epilep-tic seizures and seizure-free patients was implemented usingMATLAB software A TQWT comparison between theseizure-free and the epileptic seizures patients are shown inFigure 3 with 119876 = 1 119903 = 3 and 119869 = 3 It can be observedthat both of the amplitudes and the frequency of the epilepticseizure are much higher than the healthy one Moreover theoscillatory behavior of the epileptic patient is higher than thehealthy one The value of the parameter 119903 is set to three toprevent any excessive ringing of the wavelet as suggested by[22] MATLAB for the TQWT toolbox is available for publicaccess at httpeewebpolyeduiselesniTQWT
After the construction of the dataset the feature reduc-tion was applied using the firefly algorithm to remove theredundant and irrelevant features The number of the datasegments was determined by the parameters 119876 119903 and 119869 Bytuning these parameters the number of the data segmentswas varied and thus the training process was adopted Thetrial and error approach was used to set the value of theseparameters The experiment was implemented on variousvalues of J to obtain the best level of decomposition The
performance measures were calculated for each level ofdecomposition As shown in Figure 4 the best value of thevariable 119869 was from two to three Moreover the best value ofthe parameter119876was found to be one as shown in Figure 5 Allthe variables remained constant while changing the Q-factorto obtain the best value
Then the firefly algorithm with a population size of 20fireflies mutation probability of 001 and light absorptionequal to 01 was applied to reduce the feature setThe result ofthe compact set of features is obtained after 100 iterations andshown in Figure 6 are the number of features obtained by thefirefly algorithm and their corresponding accuracy precisionspecificity and the recall In the last step the feature set isfed to a random forest classifier to obtain the seizure-free andseizure conditionsThe random forest classifier obtained 98accuracy 97 precision 97 specificity 98 recall 98 F-measure and 95 MCC at the third level of decompositionand the Q-factor equals 1
The firefly algorithm reduced the search space into threefeatures These features could replace the original datasetwhich minimizes the processing time The compact set offeature contains the 119878119879119863 119860119901119864119899 and the 119870119860119879119885 the classi-fication rules are based on these features The classificationrules obtained by the proposed system are shown in Figure 7where the decision tree consists of 4 leaves and 7 nodes andthe final decision represents the seizure-free and the epilepticseizure denoted by 01 respectively
A comparative study of the proposed hybrid epilepsydetection approach and other existing classification systemshas been performed in terms of the total accuracy A novelmethod based on the EMDs was proposed to detect theepileptic seizures of epilepsy This method used the Hilberttransformation of IMFs obtained by EMD process thatprovided an analytic signal representation of IMFs [12] Theclassification rules obtained by this method achieved anaccuracy of 90 The usage of frequency domain featuresand Burgrsquos method obtained 9311 accuracy with SVMclassifier [43] Nonlinear features have been used with aGaussianmixturemodel classifier and achieved 95accuracy[44] A decision tree classifier is used with energy fractaldimension and sample entropy and provided 957 [45]A combination of ApEn and the Hurst exponent has beenused to detect the diagnostics of epilepsy and produced 965accuracy with SVM classifier and ANN [46] In [47] aneigensystem based method was proposed cooperated withMultiple Layer Perceptron to classify the epileptic seizureshealthy and the seizure-free This approach provided anaverage accuracy of 975 The EMD methods for epilepsydetection achieved 9775 accuracy [6]The Kraskov entropyis also combinedwith the SVMclassifier and provided 9775[22] An automated diagnostics system based on a set ofentropies and fuzzy Sugeno classifier (FSC) achieved accu-racy up to 98 [19] The work presented by [48] developeda method for the epilepsy detection using the EMD Thegenerated IMFs using the EMD were represented as a set ofamplitude and frequency modulated (AMndashFM) signals Thetwo bandwidths namely amplitude modulation bandwidthand frequency modulation bandwidth calculated from theanalytic IMFs have been fed to LS-SVM for classifying
International Journal of Biomedical Imaging 9
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(a) EEG Sample for patient with seizure-free
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(b) EEG Sample for patient with seizures
Figure 3 TQWT decomposition of seizure-free and seizure EEG signals with 119876 = 1 119903 = 3 119895 = 3
1 2 3 4 5 10LEVELS OF DECMPOSISTION (J)
ACCPrecision
SPECRecall
000
020
040
060
080
100
PERC
ENTA
GE
Figure 4 Performance measures of the proposed approach withvaried levels of decomposition
08082084086088
09092094096098
1
1 2 3 5 7 9 10
PERC
ENTA
GE
Q
ACCPrecision
SPECRecall
Figure 5 Performance measures of the proposed approach withvaried levels of decomposition
seizure and nonseizure EEG signals This method achieved9818 average accuracy The LBP-based methods have beencombined with Gabor filter for texture extraction from theEEG signal A k-nearest neighbor classifier was applied andobtained an accuracy of 983 [14] The proposed methodconfirmed its superiority in the total accuracy compared tothe other systems The results which prove the superiority ofthe proposed method compared to the other existed systemsis demonstrated in Figure 8
ACCPrecision
SPECRecall
09091092093094095096097098099
1
1 2 3 4 5PE
RCEN
TAG
ENUMBER OF FEATURES
Figure 6 Performance measures of the original and compactfeatures set
STD
lt= 2873 gt 2873
0
0
ApEn
lt= 016 gt 016
Katz 1
1
lt= 146 gt 146
Figure 7 The classification rules obtained from the proposedsystem to classify the seizure-free (0) and the epileptic seizure (1)signals
Once the system detects the preictal phase the clinicreceives a notification about that patientThemain advantageof the hybrid system from the clinical point of view could besummarized as follows
(i) Classification and detecting of the epileptic seizuresand seizure-free signals from the EEG signal automat-ically
10 International Journal of Biomedical Imaging
90
931195 957 965
975 9775 9775 98 9813 983 99
Baja
j et a
l 20
13
Faus
t et a
l 20
10
Acha
rya e
t al
2009
Mar
tis et
al 2
013
Yuan
et al
201
1
Nag
hsh-
Nilc
hi et
al 2
010
Pach
ori e
t al
2014
Patid
ar et
al 2
017
Acha
rya e
t al
2012
Baja
j et a
l 20
12
Kum
ar et
al 2
015
e p
ropo
sed
met
hod
PROPOSED SYSTESMS USED TO FOR THE COMPARSION
8486889092949698
100
ACCU
RACY
IN P
ERCE
NTA
GE
()
Figure 8 The total accuracy () of various methods employed to detect the disorder of the epilepsy
(ii) The detection of the preictal phase from studying thehealthy and ictal phase of each patient
(iii) The detection of the preictal phase that provided theability to send warning alert to the physicians toprepare the medical assessment for the patient
(iv) The proposed method being robust and reliable as itsperformance was benchmarked using 10-fold cross-validation
(v) The dynamic behavior obtained from the fireflyalgorithm which made the system adaptive to manyfeatures according to each case study
(vi) A few sets of parameters required to analyze the EEGsignal typical three variables (119876 119903 119869)
The limitations of this research were summarized as follows
(i) The limited number of the studied subjects (typically100 per class)
(ii) The diagnosis process that may be reduced because ofdepending on some additional software prerequisites
5 Conclusion
In this paper an automated intelligent CAD tool has beenproposed to classify and detect epileptic seizures and seizure-free EEG signals This method provided an EEG signalanalysis using a hybrid data fusion method The data fusionmethod combined the collected features from two differentperspectives In the first perspective the EEG signal wasconsidered as an image Then the image was converted toa gray image and the GLCM was obtained to extract thetextures of the image such as contrast correlation power andhomogeneity In the second perspective the EEG signal wasdivided into smaller segments using the TQWT to extractthe time and frequency features A set of statistical nonlinear(chaotic) and power spectrum features were obtained from
each segment After the dataset was constructed a featurereduction algorithmbased on firefly optimizationwas used toreduce the irrelevant features and remove redundancy Thenan RFwas trained to classify and predict the epileptic seizuresand seizure-free EEG signals from the dataset The experi-mental results showed that the proposed method achieveda satisfactory degree of 99 accuracy 97 precision 97specificity 98 recall 98 F-measure and 95 MCC at thethird level of decomposition
Data Availability
The dataset used to support the findings of this research wasincluded in the article and can be found as a citation to ldquoRGAndrzejak K Lehnertz F Mormann C Rieke P DavidCE Elger Indications ofNonlinearDeterministic and Finite-Dimensional Structures in Time Series of Brain ElectricalActivity Dependence on Recording Region and Brain StatePhysical Review E 64 (2001) 061907rdquo
Conflicts of Interest
The authors declare that they have no conflicts of interest
References
[1] Epilepsy Fact Sheet 2018 httpwwwwhointmediacentrefactsheetsfs999en
[2] U R Acharya S V Sree G Swapna R J Martis and J S SurildquoAutomated EEG analysis of epilepsy a reviewrdquo Knowledge-Based Systems vol 45 pp 147ndash165 2013
[3] A Puce and M S Hamalainen ldquoA review of issues relatedto data acquisition and analysis in EEGMEG studiesrdquo BrainSciences vol 7 no 6 2017
[4] V Joshi R B Pachori and A Vijesh ldquoClassification of ictaland seizure-free EEG signals using fractional linear predictionrdquoBiomedical Signal Processing and Control vol 9 pp 1ndash5 2014
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
4 International Journal of Biomedical Imaging
The final set of features consists of statistical measuresof an image represented as matrices called gray-level co-occurrence matrix (GLCM) where 119862(119894 119895) represents an entryin co-occurrence matrix and 119894 119895 = 0 1 2 119871 minus 1 where119871 is the number of gray levels in the image Those matricesrepresent the spatial dependencies between the gray levels ofimage reflecting the structure of the underlying texture Afterthe normalization of these matrices the contrast correlationenergy and homogeneity are computed as follows
119864119899119890119903119892119910 = 119871minus1sum119894=0
119871minus1sum119895=0
119862 (119894 119895)2 (18)
119862119900119899119905119903119886119904119905 = 119871minus1sum119894=0
119871minus1sum119895=0
(119894 minus 119895)2 119862 (119894 119895) (19)
119862119900119903119903119890119897119886119905119894119900119899 = 119871minus1sum119894=0
119871minus1sum119895=0
(119894 minus 120583119894) (119895 minus 120583119895) 119862 (119894 119895)120590119894120590119895 (20)
119871119900119888119886119897 ℎ119900119898119900119892119890119899119890119894119905119910 = 119871minus1sum119894=0
119871minus1sum119895=0
11 + (119894 minus 119895)2119862 (119894 119895) (21)
23 Firefly Optimization Algorithm The firefly algorithm isa swarm based stochastic search technique [37] The fireflyoptimization algorithm consists of a set of members calledfireflies each firefly represents a candidate solution Themost attractive firefly is considered to be the leader fireflythat leads the other candidates to the best region Theattractiveness is calculated based on the light intensity whichis usually determined by the objective fitness function Theattractiveness between two fireflies 119883119894 and 119883119895 is determinedas follows
120573 (119903119894119895) = 1205730119890minus1205741199032119894119895 (22)
119903119894119895 = radic 119863sum119889=1
(119909119894119889 minus 119909119895119889)2 (23)
where 119863 denotes the problem dimension such that 119863 =1 2 119889 119903119894119895 denotes the distance between 119883119894 and 119883119895Parameter 1205730 denotes the initial attractiveness at 119903 = 0 and120574 denotes the light absorption factor such that 120574 isin [0 1]Each firefly 119883119894 is compared with the other fireflies 119883119895 where119895 isin 1 2 119873 such that 119894 = 119895 and119873 denotes the count of thefireflies If firefly 119883119894 is better (brighter) than 119883119895 then firefly119883119895 moves towards 119883119894 with a step movement formulated asfollows
119883119894119889 (119905 + 1) = 119883119894119889 (119905) + 1205730119890minus1205741199032119894119895119894119895 (119883119895119889 (119905) minus 119883119894119889 (119905))+ 120572120598119894 (24)
where 120598119894 represents uniform a randomly distributed variablesuch that 120598119894 isin [minus05 05] and 120572 denotes the movement stepsuch that 120572 isin [0 1]
3 The Combinational Hybrid System ofEpilepsy Detection from EEG Signal
In this research a hybrid system was proposed to detectboth seizures and seizure-free conditions from a raw EEGsignal Although some investigations focused on the featureextraction level the proposed system was established basedon four main levels This system combined the data fusionapproach with firefly optimization and random forest The119879119876119882119879 was applied for EEG signal decomposition then thefeatures were constructed using a data fusion technique Dueto the large number of features obtained for each subband(119891119890119886119905119906119903119890119904119862119900119906119899119905 times 119869 times 119876) a feature reduction was applied toreduce the features and to obtain a compact set of featuresinstead of the original one The obtained compact set offeatures was fed to a random forest algorithm to obtainthe classification rules and hence used for training Aftertraining the classifier should be able to classify and estimatethe preictal phase The proposed system was divided intothe following four levels of processing and then described indetail as shown in Figure 1
(i) EEG decomposition using TQWT(ii) Feature extraction using data fusion based on single-
channel EEG signal and co-occurrence matrix(iii) Feature reduction using firefly optimization algo-
rithm(iv) Training of random forest classifier to detect the
seizures and seizure-free EEGs
31 EEG Decomposition Using TQWT The preprocessinglevel applies the 119879119876119882119879 decomposition to the input EEGsignal The 119879119876119882119879 converted the continuous EEG signalto discrete potions of data that could be handled moreeffectively This wavelet transformation is used because of itseffectiveness in signal decomposition and its tunability Theobtained subbands using the 119879119876119882119879 provided a significantdifference between the seizure-free and the epileptic seizureof the EEG signals as shown in Figure 2 The subfiguresdenoted by (a) (b) (c) and (d) visualize the histogram of thefirst second third and fourth subbands of the seizure-freeclass The remaining subfigures denoted by (e) (f) (g) and(h) represent the histogram of the epileptic seizure class forthe same subbands The values of the extracted subbands ofthe second class are about ten times stronger than the firstclass that prove the efficiency of this decomposition
32 Feature Extraction Using Data Fusion Based on a Single-Channel EEG and a Co-Occurrence Matrix In the firstperspective the EEG signal was described as a nonstationarytime series After the decomposition of the EEG signal usingthe TQWT a feature extraction process was performed toobtain significant characteristics fromeachTQWTsubbandsThe extracted features were categorized into three maingroups The first group determines the statistical character-istics in the time domain The mean standard deviationvariance Shannon entropy and approximate entropy werecalculated in the first group as formulated from (5) to (9)
International Journal of Biomedical Imaging 5
Image Extract
chaotic features Apply TQWT for the signal decomposition Train the random forest
classifier
EEG Signal
Extractstatistical features Feature reduction using
firefly algorithm End Constrcut feature space using data
fusion
Classifiy and estimate seizures using the trained
model
Single EEG Channel
Extract power spectrum
features
Extract Co-occurrence
matrix
Input Level 1 Level 2 Level 3 Level 4
Figure 1 Block diagram of the combinational hybrid system of epilepsy detection from EEG signal with 4 levels of processing
2500
2000
1500
1000
500
0
occu
ranc
e
minus20 0 20
V
(a)
0 20 40minus20minus40
V
0
500
1000
1500
2000
2500
3000
occu
ranc
e
(b)
0
200
400
600
800
1000
1200
1400
0 50 100minus50
V
occu
ranc
e
(c)
0
200
400
600
800
1000
1200
0 200 400minus200
V
occu
ranc
e
(d)
0
500
1000
1500
2000
2500
0 200minus200
V
occu
ranc
e
(e)
2500
2000
1500
1000
500
00 200 400minus200minus400
V
occu
ranc
e
(f)
0 500minus500
V
0
200
400
600
800
1000
1200
occu
ranc
e
(g)
0
200
400
600
800
1000
1200
0 500 1000minus500minus1000
V
occu
ranc
e
(h)
Figure 2 Illustration of seizure-free and epileptic seizures of subbands obtained from TQWT with Q = 1 r = 3 J = 3 Figures (a) (b) (c)and (d) represent the first second third and fourth subbands obtained from seizure-free signals Figures (e) (f) (g) and (h) represent thefirst second third and fourth subbands obtained from the epileptic seizures
This group indicates some statistical information obtainedfrom the time domain of TQWT subband The secondgroup of features determines a power spectrum analysisof obtained subbands The discrete Fourier transformations(DFT) were applied to convert these subbands into a fre-quency domain Then the power spectrum features wereextractedThe second group consists of the spectral centroidspectral speed spectral flatness spectral slope and spectralentropy as formulated from (10) to (14) The power spectrumanalysis represents an effectivemethod to study the frequencybehavior of the signal The last group of features performs achaotic analysis of each subband Because of the nonlinearityof the EEG signals a nonlinear analysis is required One ofthe best analyses used for this issue is the chaotic analysisIn this analysis the Higuchi fractal dimension (HFD) Hurst
exponent and Katz fractal exponent were computed for eachsubband as formulated from (15) to (18)
In the second perspective the input EEG signal wasconverted to a gray image Then a co-occurrence matrix wascomputed to obtain the gray levels of the image Then thetextures of contrast correlation energy and homogeneitywere calculated from this matrix to represent the statisticalmeasures of the image as formulated from (19) to (21) Finallyafter computing the feature space a data fusion was appliedto merge all of these features and create a single dataset
33 Feature Reduction Using Firefly Optimization AlgorithmIn this section a feature reduction algorithm based onthe firefly algorithm is proposed [37 38] This algorithmimplements a chaotic movement simulated annealing (SA)
6 International Journal of Biomedical Imaging
(1) InputThe features matrix 119865119898(2) OutputThe optimal solution 119892119887119890119904119905(3) Initialize the firefly swam using a logistic chaotic map(4) Evaluate each firefly 119892119894 using the fitness function f (x)(5) Select the best firefly 119892119887119890119904119905 and the worst one 119892119908119900119903119904119905(6) while termination condition are not reached do(7) Declare an alternative leader firefly as 119878119887119890119904119905 with a competitive fitness and located in different region(8) Obtain an offspring solution 119892119887119890119904119905 using 119878119860(9) for all (firefly 119894 and 119891119894 = 119892119908119900119904119903119905) do(10) for all (firefly 119895 and 119891119895 = 119892119908119900119904119903119905) do(11) if (119868119895 gt 119868119894) then(12) Enhance firefly 119895 using Equation (25) to obtain better offspring candidate firefly denoted 1199091015840119895(13) Replace firefly 119895 with the offspring 1199091015840119895(14) Move the firefly 119895 towards the neighboring and global optimal solutions using Equation (26)(15) end if(16) end for(17) end for(18) Update the worst solution 119892119908119900119903119904119905(19) if 119891(1198921015840119887119890119904119905) gt 119891(119892119887119890119904119905) then(20) 119892119887119890119904119905 larr997888 1198921015840119887119890119904119905(21) end if(22) Rank all fireflies and update the best 119892119887119890119904119905 and worst 119892119908119900119903119904119905 solutions(23) end while(24) return 119892119887119890119904119905Algorithm 1 The pseudocode of the proposed feature reduction algorithm based on firefly optimization and SA
to produce efficient offspring candidates andmemory aware-ness of the best and worst solutions to improve the searchdiversity and prevent local optima The firefly population israndomly initialized using a chaotic logistic map to ensurethe diversely of the candidates and the randomness of eachfirefly Afterwards the fitness function is computed for eachcandidate and identifies both the best and worst solutions as119892119887119890119904119905 and 119892119908119900119904119903119905 respectively For each iteration an alternativecandidate is declared as 119878119887119890119904119905 with a competitive fitnessand located in a different region Both of the best and thealternative candidate sets are used to lead weak solutions toreach the optimal region and prevent local optima problemThe mean of the leader firefly and the alternative one isenhanced using SA algorithm to obtain a better solution1198921015840119887119890119904119905 The improved local and global solutions are used toguide the low lightness fireflies to move towards that strongerlightness The algorithm is repeated until the maximumnumber of iterations is reached or a termination criterion isachieved The behavior of the proposed algorithm is deter-mined by some properties namely the objective function theattractiveness movement step and population diversity Theproposed algorithm is demonstrated in Algorithm 1
TheObjective FunctionThis function is used to evaluate eachcandidate in the algorithm and defined as follows
119891 (119909) = 1199081 times 119886119888119888119906119903119886119888119910 (119909) + 1199082number of features (25)
where 1199081 and 1199082 represent the weights of the classifi-cation accuracy and the number of the selected features
respectively The values of 1199081 and 1199082 are set to 09 and 01respectively as a recommended by [38]
The Attractiveness Movement Step The proposed algorithmused a chaotic logistic map to initialize the firefly populationand hence increases the diversity and avoid local optimaAfter obtaining the global best solution 119892119887119890119904119905 an alternativeleader firefly 119878119887119890119904119905 is declared with a competitive fitness butlocated in a different region Since both leaders are morelikely to discover distinctive search regions this strategyreduces the probability of being trapped in the local optimaIn addition the optimal offspring of themean positions of thetwo leaders and the neighboring brighter candidates are usedto lead the search process and guide the solutions with lowerlight intensity to move towards the optimal region
119909119894 = 119909119894 + 1205730119862119896 (1199091015840119895 minus 119909119894) + 119862119896120576 (1198921015840119887119890119904119905 minus 119909119894) + 1205721015840times sign [119903119886119899119889 minus 05] (26)
1199091015840119895 = 119909119895 + 1205901 (27)
1198921015840119887119890119904119905 = 119898119890119886119899 (119892119887119890119904119905 + 119878119887119890119904119905) + 1205902 (28)
where 1199091015840119895 denotes the offspring candidate with a brighterneighboring solution119909119895 is defined by the SA as shown in (26)and 1198921015840119887119890119904119905 represents the fitter offspring solutions of the meanof the leader firefly and the alternative one as formulated in(27) It worth mentioning that the values of 1205901 and 1205902 aretwo random variables set using the Gaussian distributionThe movement step of the firefly is determined as shown in
International Journal of Biomedical Imaging 7
(1) Input 119894119905119890119903119886119905119894119900119899119904119898119886119909 119892119898119890119886119899 119879119898119886119909(2) OutputThe optimal offspring 119878119887119890119904119905(3) 119878119888 = 119862119903119890119886119905119890119868119899119894119905119878119900119906119897119905119894119900119899119904(119892119898119890119886119899)(4) 119878119887119890119904119905 = 119878119888(5) while (119894 le 119894119905119890119903119886119905119894119900119899119904 119898119886119909) do(6) 119878119894 = 119862119903119890119886119905119890119873119890119894119892ℎ119887119900119903119878119900119897119906119905119894119900119899(119878119888)(7) 119879119888 = 119862119886119897119888119906119897119886119905119890119879119890119898119901119890119903119886119905119906119903119890(119894 119879119898119886119909)(8) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119888)) then(9) 119878119888 = 119878119894(10) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119887119890119904119905)) then(11) 119878119887119890119904119905 = 119878119894(12) end if(13) else if (exp((119888119900119904119905(119878119887119890119904119905) minus 119888119900119904119905(119878119894))119879119888) ge 120590) then(14) 119878119888 = 119878119894(15) end if(16) 119894 = 119894 + 1(17) end while(18) return 119878119887119890119904119905Algorithm 2 The pseudocode of the generation of offspring solutions using SA
(28) where 119862119870 represents the chaotic map variable in themovement step and 120576 denotes the randomized vector definedin the traditional firefly algorithm Parameter 1205721015840 denotes anadaptive step initialized to 05 to control the diversity of thesearch process
Offspring Generation Using SAThe proposed algorithm usedSA for generating better candidates to enhance the searchprocess as much as possible as shown in Algorithm 2 TheSA accepts both of the best solution 119892119887119890119904119905 and the alternativesolution 119878119887119890119904119905 as main inputs then the traditional SA isapplied The better solution generated is accepted by defaultaccording to the SA heuristics On the other hand theweaker solution should be accepted with specific probabilityas shown in (29) where Δ119891 denotes the difference of thefitness (energy) between to candidates and 119879119888 denotes thecurrent temperature A simple linear cooling mechanism isused to control the value of the temperature
119901119909119895 = exp(Δ119891119879119888 ) (29)
Population Diversity In each iteration the worst solution isdetected after the ranking process as shown in Algorithm 1The remaining solutions are guided by the average positionobtained from (30) where 120590 denotes a random value obtainedby Gaussian map
119909119908119900119903119904119905119895 = 119892119887119890119904119905 + 1198781198871198901199041199052 + 120590 (119909119908119900119903119904119905119895 119892119887119890119904119905 + 1198781198871198901199041199052 ) (30)
34 Classification and Learning The random forest (RF) isa successful ensemble approach used in supervised machinelearning to solve classification or regression problems [39]It consists of a collection of decision trees that could actas a single classifier with multiple classification methodsor a method that has several variables Several subsets ofthe training data are supplied to each tree to achieve the
most stable tree classification that results in a generalizedexperience of the classifier The original dataset is dividedinto two parts The first part is used to train each tree bybootstrapping technique The other part is used to evaluatethe accuracy of the classification Each tree is allowed to reachthe maximum depth without tree pruning to obtain a highvariance classifier The splitting process remains until onlyone instance of a single class is dropped from any leaf nodeor a predefined termination condition is achieved Whenthe forest is established the number of subsets remained asa constant The obtained route of traversal from the rootnode to the leaf node is applied to the new instances orthe unlabeled instance for classification The final decisionfor classifying a new instance is provided by determiningeach class that has the most votes from every decision treeThe random forest performs slightly better when comparedwith other classifiers such as discriminant analysis SVM andartificial neural network (ANN) [15 39]
35 Performance Evaluation Various performance formulaswere used to evaluate the effectiveness of a classifierThe sen-sitivity (SEN) or recall specificity (SPEC) accuracy (ACC)F-measure Matthewrsquos correlation coefficient (MCC) andreceiver operating characteristics (ROC) are used to evaluatethe efficiency of the random forest classifier [40ndash42] Theseparameters are defined as follows
119878119864119873 = 119879119875119879119875 + 119865119873 times 100 (31)
119878119875119864119862 = 119879119873119879119873 + 119865119875 times 100 (32)
119860119862119862 = 119879119875 + 119879119873119879119875 + 119879119873 + 119865119875 + 119865119873 times 100 (33)
119865 minus 119898119890119886119904119906119903119890 = 21198791198752119879119875 + 119865119875 + 119865119873 times 100 (34)
8 International Journal of Biomedical Imaging
119872119862119862= (119879119875 times 119879119873) minus (119865119875 times 119865119873)radic(119879119875 + 119865119873) (119879119875 + 119865119875) (119879119873 + 119865119873) (119879119873 + 119865119875)times 100
(35)
119879119875 and 119879119873 represent the total number of an epileptic seizureand seizure-free signals classified correctly respectively Sim-ilarly 119865119875 and 119865119873 represent the total number of epilepticseizures and seizure-free signals classified incorrectly respec-tively Cross-validation has also been used to ensure theclassifier reliability and effectiveness The original dataset issplit into 119896 folds (subsets) for both training and testing Inthis strategy 119896 minus 1 folds were selected randomly to train theclassifier and the remaining folds are used to testing Theoverall performance is calculated as the average of each foldIn this work the experiment is repeated 10 times with tenfoldcross-validation
4 Results and Discussion
The benchmark dataset used in this investigation wasacquired by the University of Bonn [28]The dataset containsthree different categories ie preictal healthy and ictalrecorded using a single channel for a 236 s duration Bothnormal and preictal conditions were collected from 200 casestudies and 100 for the ictal state The normal condition isacquired from five healthy volunteers using the international10ndash20 system standard with each volunteer in a relaxed-awake state with eyes open and closed (100 cases per each set)[28] The ictal data were collected from five patients duringtheir epileptic seizures The preictal represents the EEG datacollected from the same five patients with no seizures It isworth mentioning that all EEG signals were acquired using a128 channel amplifier with sampling rate equal 17361Hz [28]Finally a bandpass filter with 05340Hz sim12 dBoctave wasapplied as a filter
The proposed approach for automatic detection of epilep-tic seizures and seizure-free patients was implemented usingMATLAB software A TQWT comparison between theseizure-free and the epileptic seizures patients are shown inFigure 3 with 119876 = 1 119903 = 3 and 119869 = 3 It can be observedthat both of the amplitudes and the frequency of the epilepticseizure are much higher than the healthy one Moreover theoscillatory behavior of the epileptic patient is higher than thehealthy one The value of the parameter 119903 is set to three toprevent any excessive ringing of the wavelet as suggested by[22] MATLAB for the TQWT toolbox is available for publicaccess at httpeewebpolyeduiselesniTQWT
After the construction of the dataset the feature reduc-tion was applied using the firefly algorithm to remove theredundant and irrelevant features The number of the datasegments was determined by the parameters 119876 119903 and 119869 Bytuning these parameters the number of the data segmentswas varied and thus the training process was adopted Thetrial and error approach was used to set the value of theseparameters The experiment was implemented on variousvalues of J to obtain the best level of decomposition The
performance measures were calculated for each level ofdecomposition As shown in Figure 4 the best value of thevariable 119869 was from two to three Moreover the best value ofthe parameter119876was found to be one as shown in Figure 5 Allthe variables remained constant while changing the Q-factorto obtain the best value
Then the firefly algorithm with a population size of 20fireflies mutation probability of 001 and light absorptionequal to 01 was applied to reduce the feature setThe result ofthe compact set of features is obtained after 100 iterations andshown in Figure 6 are the number of features obtained by thefirefly algorithm and their corresponding accuracy precisionspecificity and the recall In the last step the feature set isfed to a random forest classifier to obtain the seizure-free andseizure conditionsThe random forest classifier obtained 98accuracy 97 precision 97 specificity 98 recall 98 F-measure and 95 MCC at the third level of decompositionand the Q-factor equals 1
The firefly algorithm reduced the search space into threefeatures These features could replace the original datasetwhich minimizes the processing time The compact set offeature contains the 119878119879119863 119860119901119864119899 and the 119870119860119879119885 the classi-fication rules are based on these features The classificationrules obtained by the proposed system are shown in Figure 7where the decision tree consists of 4 leaves and 7 nodes andthe final decision represents the seizure-free and the epilepticseizure denoted by 01 respectively
A comparative study of the proposed hybrid epilepsydetection approach and other existing classification systemshas been performed in terms of the total accuracy A novelmethod based on the EMDs was proposed to detect theepileptic seizures of epilepsy This method used the Hilberttransformation of IMFs obtained by EMD process thatprovided an analytic signal representation of IMFs [12] Theclassification rules obtained by this method achieved anaccuracy of 90 The usage of frequency domain featuresand Burgrsquos method obtained 9311 accuracy with SVMclassifier [43] Nonlinear features have been used with aGaussianmixturemodel classifier and achieved 95accuracy[44] A decision tree classifier is used with energy fractaldimension and sample entropy and provided 957 [45]A combination of ApEn and the Hurst exponent has beenused to detect the diagnostics of epilepsy and produced 965accuracy with SVM classifier and ANN [46] In [47] aneigensystem based method was proposed cooperated withMultiple Layer Perceptron to classify the epileptic seizureshealthy and the seizure-free This approach provided anaverage accuracy of 975 The EMD methods for epilepsydetection achieved 9775 accuracy [6]The Kraskov entropyis also combinedwith the SVMclassifier and provided 9775[22] An automated diagnostics system based on a set ofentropies and fuzzy Sugeno classifier (FSC) achieved accu-racy up to 98 [19] The work presented by [48] developeda method for the epilepsy detection using the EMD Thegenerated IMFs using the EMD were represented as a set ofamplitude and frequency modulated (AMndashFM) signals Thetwo bandwidths namely amplitude modulation bandwidthand frequency modulation bandwidth calculated from theanalytic IMFs have been fed to LS-SVM for classifying
International Journal of Biomedical Imaging 9
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(a) EEG Sample for patient with seizure-free
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(b) EEG Sample for patient with seizures
Figure 3 TQWT decomposition of seizure-free and seizure EEG signals with 119876 = 1 119903 = 3 119895 = 3
1 2 3 4 5 10LEVELS OF DECMPOSISTION (J)
ACCPrecision
SPECRecall
000
020
040
060
080
100
PERC
ENTA
GE
Figure 4 Performance measures of the proposed approach withvaried levels of decomposition
08082084086088
09092094096098
1
1 2 3 5 7 9 10
PERC
ENTA
GE
Q
ACCPrecision
SPECRecall
Figure 5 Performance measures of the proposed approach withvaried levels of decomposition
seizure and nonseizure EEG signals This method achieved9818 average accuracy The LBP-based methods have beencombined with Gabor filter for texture extraction from theEEG signal A k-nearest neighbor classifier was applied andobtained an accuracy of 983 [14] The proposed methodconfirmed its superiority in the total accuracy compared tothe other systems The results which prove the superiority ofthe proposed method compared to the other existed systemsis demonstrated in Figure 8
ACCPrecision
SPECRecall
09091092093094095096097098099
1
1 2 3 4 5PE
RCEN
TAG
ENUMBER OF FEATURES
Figure 6 Performance measures of the original and compactfeatures set
STD
lt= 2873 gt 2873
0
0
ApEn
lt= 016 gt 016
Katz 1
1
lt= 146 gt 146
Figure 7 The classification rules obtained from the proposedsystem to classify the seizure-free (0) and the epileptic seizure (1)signals
Once the system detects the preictal phase the clinicreceives a notification about that patientThemain advantageof the hybrid system from the clinical point of view could besummarized as follows
(i) Classification and detecting of the epileptic seizuresand seizure-free signals from the EEG signal automat-ically
10 International Journal of Biomedical Imaging
90
931195 957 965
975 9775 9775 98 9813 983 99
Baja
j et a
l 20
13
Faus
t et a
l 20
10
Acha
rya e
t al
2009
Mar
tis et
al 2
013
Yuan
et al
201
1
Nag
hsh-
Nilc
hi et
al 2
010
Pach
ori e
t al
2014
Patid
ar et
al 2
017
Acha
rya e
t al
2012
Baja
j et a
l 20
12
Kum
ar et
al 2
015
e p
ropo
sed
met
hod
PROPOSED SYSTESMS USED TO FOR THE COMPARSION
8486889092949698
100
ACCU
RACY
IN P
ERCE
NTA
GE
()
Figure 8 The total accuracy () of various methods employed to detect the disorder of the epilepsy
(ii) The detection of the preictal phase from studying thehealthy and ictal phase of each patient
(iii) The detection of the preictal phase that provided theability to send warning alert to the physicians toprepare the medical assessment for the patient
(iv) The proposed method being robust and reliable as itsperformance was benchmarked using 10-fold cross-validation
(v) The dynamic behavior obtained from the fireflyalgorithm which made the system adaptive to manyfeatures according to each case study
(vi) A few sets of parameters required to analyze the EEGsignal typical three variables (119876 119903 119869)
The limitations of this research were summarized as follows
(i) The limited number of the studied subjects (typically100 per class)
(ii) The diagnosis process that may be reduced because ofdepending on some additional software prerequisites
5 Conclusion
In this paper an automated intelligent CAD tool has beenproposed to classify and detect epileptic seizures and seizure-free EEG signals This method provided an EEG signalanalysis using a hybrid data fusion method The data fusionmethod combined the collected features from two differentperspectives In the first perspective the EEG signal wasconsidered as an image Then the image was converted toa gray image and the GLCM was obtained to extract thetextures of the image such as contrast correlation power andhomogeneity In the second perspective the EEG signal wasdivided into smaller segments using the TQWT to extractthe time and frequency features A set of statistical nonlinear(chaotic) and power spectrum features were obtained from
each segment After the dataset was constructed a featurereduction algorithmbased on firefly optimizationwas used toreduce the irrelevant features and remove redundancy Thenan RFwas trained to classify and predict the epileptic seizuresand seizure-free EEG signals from the dataset The experi-mental results showed that the proposed method achieveda satisfactory degree of 99 accuracy 97 precision 97specificity 98 recall 98 F-measure and 95 MCC at thethird level of decomposition
Data Availability
The dataset used to support the findings of this research wasincluded in the article and can be found as a citation to ldquoRGAndrzejak K Lehnertz F Mormann C Rieke P DavidCE Elger Indications ofNonlinearDeterministic and Finite-Dimensional Structures in Time Series of Brain ElectricalActivity Dependence on Recording Region and Brain StatePhysical Review E 64 (2001) 061907rdquo
Conflicts of Interest
The authors declare that they have no conflicts of interest
References
[1] Epilepsy Fact Sheet 2018 httpwwwwhointmediacentrefactsheetsfs999en
[2] U R Acharya S V Sree G Swapna R J Martis and J S SurildquoAutomated EEG analysis of epilepsy a reviewrdquo Knowledge-Based Systems vol 45 pp 147ndash165 2013
[3] A Puce and M S Hamalainen ldquoA review of issues relatedto data acquisition and analysis in EEGMEG studiesrdquo BrainSciences vol 7 no 6 2017
[4] V Joshi R B Pachori and A Vijesh ldquoClassification of ictaland seizure-free EEG signals using fractional linear predictionrdquoBiomedical Signal Processing and Control vol 9 pp 1ndash5 2014
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
International Journal of Biomedical Imaging 5
Image Extract
chaotic features Apply TQWT for the signal decomposition Train the random forest
classifier
EEG Signal
Extractstatistical features Feature reduction using
firefly algorithm End Constrcut feature space using data
fusion
Classifiy and estimate seizures using the trained
model
Single EEG Channel
Extract power spectrum
features
Extract Co-occurrence
matrix
Input Level 1 Level 2 Level 3 Level 4
Figure 1 Block diagram of the combinational hybrid system of epilepsy detection from EEG signal with 4 levels of processing
2500
2000
1500
1000
500
0
occu
ranc
e
minus20 0 20
V
(a)
0 20 40minus20minus40
V
0
500
1000
1500
2000
2500
3000
occu
ranc
e
(b)
0
200
400
600
800
1000
1200
1400
0 50 100minus50
V
occu
ranc
e
(c)
0
200
400
600
800
1000
1200
0 200 400minus200
V
occu
ranc
e
(d)
0
500
1000
1500
2000
2500
0 200minus200
V
occu
ranc
e
(e)
2500
2000
1500
1000
500
00 200 400minus200minus400
V
occu
ranc
e
(f)
0 500minus500
V
0
200
400
600
800
1000
1200
occu
ranc
e
(g)
0
200
400
600
800
1000
1200
0 500 1000minus500minus1000
V
occu
ranc
e
(h)
Figure 2 Illustration of seizure-free and epileptic seizures of subbands obtained from TQWT with Q = 1 r = 3 J = 3 Figures (a) (b) (c)and (d) represent the first second third and fourth subbands obtained from seizure-free signals Figures (e) (f) (g) and (h) represent thefirst second third and fourth subbands obtained from the epileptic seizures
This group indicates some statistical information obtainedfrom the time domain of TQWT subband The secondgroup of features determines a power spectrum analysisof obtained subbands The discrete Fourier transformations(DFT) were applied to convert these subbands into a fre-quency domain Then the power spectrum features wereextractedThe second group consists of the spectral centroidspectral speed spectral flatness spectral slope and spectralentropy as formulated from (10) to (14) The power spectrumanalysis represents an effectivemethod to study the frequencybehavior of the signal The last group of features performs achaotic analysis of each subband Because of the nonlinearityof the EEG signals a nonlinear analysis is required One ofthe best analyses used for this issue is the chaotic analysisIn this analysis the Higuchi fractal dimension (HFD) Hurst
exponent and Katz fractal exponent were computed for eachsubband as formulated from (15) to (18)
In the second perspective the input EEG signal wasconverted to a gray image Then a co-occurrence matrix wascomputed to obtain the gray levels of the image Then thetextures of contrast correlation energy and homogeneitywere calculated from this matrix to represent the statisticalmeasures of the image as formulated from (19) to (21) Finallyafter computing the feature space a data fusion was appliedto merge all of these features and create a single dataset
33 Feature Reduction Using Firefly Optimization AlgorithmIn this section a feature reduction algorithm based onthe firefly algorithm is proposed [37 38] This algorithmimplements a chaotic movement simulated annealing (SA)
6 International Journal of Biomedical Imaging
(1) InputThe features matrix 119865119898(2) OutputThe optimal solution 119892119887119890119904119905(3) Initialize the firefly swam using a logistic chaotic map(4) Evaluate each firefly 119892119894 using the fitness function f (x)(5) Select the best firefly 119892119887119890119904119905 and the worst one 119892119908119900119903119904119905(6) while termination condition are not reached do(7) Declare an alternative leader firefly as 119878119887119890119904119905 with a competitive fitness and located in different region(8) Obtain an offspring solution 119892119887119890119904119905 using 119878119860(9) for all (firefly 119894 and 119891119894 = 119892119908119900119904119903119905) do(10) for all (firefly 119895 and 119891119895 = 119892119908119900119904119903119905) do(11) if (119868119895 gt 119868119894) then(12) Enhance firefly 119895 using Equation (25) to obtain better offspring candidate firefly denoted 1199091015840119895(13) Replace firefly 119895 with the offspring 1199091015840119895(14) Move the firefly 119895 towards the neighboring and global optimal solutions using Equation (26)(15) end if(16) end for(17) end for(18) Update the worst solution 119892119908119900119903119904119905(19) if 119891(1198921015840119887119890119904119905) gt 119891(119892119887119890119904119905) then(20) 119892119887119890119904119905 larr997888 1198921015840119887119890119904119905(21) end if(22) Rank all fireflies and update the best 119892119887119890119904119905 and worst 119892119908119900119903119904119905 solutions(23) end while(24) return 119892119887119890119904119905Algorithm 1 The pseudocode of the proposed feature reduction algorithm based on firefly optimization and SA
to produce efficient offspring candidates andmemory aware-ness of the best and worst solutions to improve the searchdiversity and prevent local optima The firefly population israndomly initialized using a chaotic logistic map to ensurethe diversely of the candidates and the randomness of eachfirefly Afterwards the fitness function is computed for eachcandidate and identifies both the best and worst solutions as119892119887119890119904119905 and 119892119908119900119904119903119905 respectively For each iteration an alternativecandidate is declared as 119878119887119890119904119905 with a competitive fitnessand located in a different region Both of the best and thealternative candidate sets are used to lead weak solutions toreach the optimal region and prevent local optima problemThe mean of the leader firefly and the alternative one isenhanced using SA algorithm to obtain a better solution1198921015840119887119890119904119905 The improved local and global solutions are used toguide the low lightness fireflies to move towards that strongerlightness The algorithm is repeated until the maximumnumber of iterations is reached or a termination criterion isachieved The behavior of the proposed algorithm is deter-mined by some properties namely the objective function theattractiveness movement step and population diversity Theproposed algorithm is demonstrated in Algorithm 1
TheObjective FunctionThis function is used to evaluate eachcandidate in the algorithm and defined as follows
119891 (119909) = 1199081 times 119886119888119888119906119903119886119888119910 (119909) + 1199082number of features (25)
where 1199081 and 1199082 represent the weights of the classifi-cation accuracy and the number of the selected features
respectively The values of 1199081 and 1199082 are set to 09 and 01respectively as a recommended by [38]
The Attractiveness Movement Step The proposed algorithmused a chaotic logistic map to initialize the firefly populationand hence increases the diversity and avoid local optimaAfter obtaining the global best solution 119892119887119890119904119905 an alternativeleader firefly 119878119887119890119904119905 is declared with a competitive fitness butlocated in a different region Since both leaders are morelikely to discover distinctive search regions this strategyreduces the probability of being trapped in the local optimaIn addition the optimal offspring of themean positions of thetwo leaders and the neighboring brighter candidates are usedto lead the search process and guide the solutions with lowerlight intensity to move towards the optimal region
119909119894 = 119909119894 + 1205730119862119896 (1199091015840119895 minus 119909119894) + 119862119896120576 (1198921015840119887119890119904119905 minus 119909119894) + 1205721015840times sign [119903119886119899119889 minus 05] (26)
1199091015840119895 = 119909119895 + 1205901 (27)
1198921015840119887119890119904119905 = 119898119890119886119899 (119892119887119890119904119905 + 119878119887119890119904119905) + 1205902 (28)
where 1199091015840119895 denotes the offspring candidate with a brighterneighboring solution119909119895 is defined by the SA as shown in (26)and 1198921015840119887119890119904119905 represents the fitter offspring solutions of the meanof the leader firefly and the alternative one as formulated in(27) It worth mentioning that the values of 1205901 and 1205902 aretwo random variables set using the Gaussian distributionThe movement step of the firefly is determined as shown in
International Journal of Biomedical Imaging 7
(1) Input 119894119905119890119903119886119905119894119900119899119904119898119886119909 119892119898119890119886119899 119879119898119886119909(2) OutputThe optimal offspring 119878119887119890119904119905(3) 119878119888 = 119862119903119890119886119905119890119868119899119894119905119878119900119906119897119905119894119900119899119904(119892119898119890119886119899)(4) 119878119887119890119904119905 = 119878119888(5) while (119894 le 119894119905119890119903119886119905119894119900119899119904 119898119886119909) do(6) 119878119894 = 119862119903119890119886119905119890119873119890119894119892ℎ119887119900119903119878119900119897119906119905119894119900119899(119878119888)(7) 119879119888 = 119862119886119897119888119906119897119886119905119890119879119890119898119901119890119903119886119905119906119903119890(119894 119879119898119886119909)(8) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119888)) then(9) 119878119888 = 119878119894(10) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119887119890119904119905)) then(11) 119878119887119890119904119905 = 119878119894(12) end if(13) else if (exp((119888119900119904119905(119878119887119890119904119905) minus 119888119900119904119905(119878119894))119879119888) ge 120590) then(14) 119878119888 = 119878119894(15) end if(16) 119894 = 119894 + 1(17) end while(18) return 119878119887119890119904119905Algorithm 2 The pseudocode of the generation of offspring solutions using SA
(28) where 119862119870 represents the chaotic map variable in themovement step and 120576 denotes the randomized vector definedin the traditional firefly algorithm Parameter 1205721015840 denotes anadaptive step initialized to 05 to control the diversity of thesearch process
Offspring Generation Using SAThe proposed algorithm usedSA for generating better candidates to enhance the searchprocess as much as possible as shown in Algorithm 2 TheSA accepts both of the best solution 119892119887119890119904119905 and the alternativesolution 119878119887119890119904119905 as main inputs then the traditional SA isapplied The better solution generated is accepted by defaultaccording to the SA heuristics On the other hand theweaker solution should be accepted with specific probabilityas shown in (29) where Δ119891 denotes the difference of thefitness (energy) between to candidates and 119879119888 denotes thecurrent temperature A simple linear cooling mechanism isused to control the value of the temperature
119901119909119895 = exp(Δ119891119879119888 ) (29)
Population Diversity In each iteration the worst solution isdetected after the ranking process as shown in Algorithm 1The remaining solutions are guided by the average positionobtained from (30) where 120590 denotes a random value obtainedby Gaussian map
119909119908119900119903119904119905119895 = 119892119887119890119904119905 + 1198781198871198901199041199052 + 120590 (119909119908119900119903119904119905119895 119892119887119890119904119905 + 1198781198871198901199041199052 ) (30)
34 Classification and Learning The random forest (RF) isa successful ensemble approach used in supervised machinelearning to solve classification or regression problems [39]It consists of a collection of decision trees that could actas a single classifier with multiple classification methodsor a method that has several variables Several subsets ofthe training data are supplied to each tree to achieve the
most stable tree classification that results in a generalizedexperience of the classifier The original dataset is dividedinto two parts The first part is used to train each tree bybootstrapping technique The other part is used to evaluatethe accuracy of the classification Each tree is allowed to reachthe maximum depth without tree pruning to obtain a highvariance classifier The splitting process remains until onlyone instance of a single class is dropped from any leaf nodeor a predefined termination condition is achieved Whenthe forest is established the number of subsets remained asa constant The obtained route of traversal from the rootnode to the leaf node is applied to the new instances orthe unlabeled instance for classification The final decisionfor classifying a new instance is provided by determiningeach class that has the most votes from every decision treeThe random forest performs slightly better when comparedwith other classifiers such as discriminant analysis SVM andartificial neural network (ANN) [15 39]
35 Performance Evaluation Various performance formulaswere used to evaluate the effectiveness of a classifierThe sen-sitivity (SEN) or recall specificity (SPEC) accuracy (ACC)F-measure Matthewrsquos correlation coefficient (MCC) andreceiver operating characteristics (ROC) are used to evaluatethe efficiency of the random forest classifier [40ndash42] Theseparameters are defined as follows
119878119864119873 = 119879119875119879119875 + 119865119873 times 100 (31)
119878119875119864119862 = 119879119873119879119873 + 119865119875 times 100 (32)
119860119862119862 = 119879119875 + 119879119873119879119875 + 119879119873 + 119865119875 + 119865119873 times 100 (33)
119865 minus 119898119890119886119904119906119903119890 = 21198791198752119879119875 + 119865119875 + 119865119873 times 100 (34)
8 International Journal of Biomedical Imaging
119872119862119862= (119879119875 times 119879119873) minus (119865119875 times 119865119873)radic(119879119875 + 119865119873) (119879119875 + 119865119875) (119879119873 + 119865119873) (119879119873 + 119865119875)times 100
(35)
119879119875 and 119879119873 represent the total number of an epileptic seizureand seizure-free signals classified correctly respectively Sim-ilarly 119865119875 and 119865119873 represent the total number of epilepticseizures and seizure-free signals classified incorrectly respec-tively Cross-validation has also been used to ensure theclassifier reliability and effectiveness The original dataset issplit into 119896 folds (subsets) for both training and testing Inthis strategy 119896 minus 1 folds were selected randomly to train theclassifier and the remaining folds are used to testing Theoverall performance is calculated as the average of each foldIn this work the experiment is repeated 10 times with tenfoldcross-validation
4 Results and Discussion
The benchmark dataset used in this investigation wasacquired by the University of Bonn [28]The dataset containsthree different categories ie preictal healthy and ictalrecorded using a single channel for a 236 s duration Bothnormal and preictal conditions were collected from 200 casestudies and 100 for the ictal state The normal condition isacquired from five healthy volunteers using the international10ndash20 system standard with each volunteer in a relaxed-awake state with eyes open and closed (100 cases per each set)[28] The ictal data were collected from five patients duringtheir epileptic seizures The preictal represents the EEG datacollected from the same five patients with no seizures It isworth mentioning that all EEG signals were acquired using a128 channel amplifier with sampling rate equal 17361Hz [28]Finally a bandpass filter with 05340Hz sim12 dBoctave wasapplied as a filter
The proposed approach for automatic detection of epilep-tic seizures and seizure-free patients was implemented usingMATLAB software A TQWT comparison between theseizure-free and the epileptic seizures patients are shown inFigure 3 with 119876 = 1 119903 = 3 and 119869 = 3 It can be observedthat both of the amplitudes and the frequency of the epilepticseizure are much higher than the healthy one Moreover theoscillatory behavior of the epileptic patient is higher than thehealthy one The value of the parameter 119903 is set to three toprevent any excessive ringing of the wavelet as suggested by[22] MATLAB for the TQWT toolbox is available for publicaccess at httpeewebpolyeduiselesniTQWT
After the construction of the dataset the feature reduc-tion was applied using the firefly algorithm to remove theredundant and irrelevant features The number of the datasegments was determined by the parameters 119876 119903 and 119869 Bytuning these parameters the number of the data segmentswas varied and thus the training process was adopted Thetrial and error approach was used to set the value of theseparameters The experiment was implemented on variousvalues of J to obtain the best level of decomposition The
performance measures were calculated for each level ofdecomposition As shown in Figure 4 the best value of thevariable 119869 was from two to three Moreover the best value ofthe parameter119876was found to be one as shown in Figure 5 Allthe variables remained constant while changing the Q-factorto obtain the best value
Then the firefly algorithm with a population size of 20fireflies mutation probability of 001 and light absorptionequal to 01 was applied to reduce the feature setThe result ofthe compact set of features is obtained after 100 iterations andshown in Figure 6 are the number of features obtained by thefirefly algorithm and their corresponding accuracy precisionspecificity and the recall In the last step the feature set isfed to a random forest classifier to obtain the seizure-free andseizure conditionsThe random forest classifier obtained 98accuracy 97 precision 97 specificity 98 recall 98 F-measure and 95 MCC at the third level of decompositionand the Q-factor equals 1
The firefly algorithm reduced the search space into threefeatures These features could replace the original datasetwhich minimizes the processing time The compact set offeature contains the 119878119879119863 119860119901119864119899 and the 119870119860119879119885 the classi-fication rules are based on these features The classificationrules obtained by the proposed system are shown in Figure 7where the decision tree consists of 4 leaves and 7 nodes andthe final decision represents the seizure-free and the epilepticseizure denoted by 01 respectively
A comparative study of the proposed hybrid epilepsydetection approach and other existing classification systemshas been performed in terms of the total accuracy A novelmethod based on the EMDs was proposed to detect theepileptic seizures of epilepsy This method used the Hilberttransformation of IMFs obtained by EMD process thatprovided an analytic signal representation of IMFs [12] Theclassification rules obtained by this method achieved anaccuracy of 90 The usage of frequency domain featuresand Burgrsquos method obtained 9311 accuracy with SVMclassifier [43] Nonlinear features have been used with aGaussianmixturemodel classifier and achieved 95accuracy[44] A decision tree classifier is used with energy fractaldimension and sample entropy and provided 957 [45]A combination of ApEn and the Hurst exponent has beenused to detect the diagnostics of epilepsy and produced 965accuracy with SVM classifier and ANN [46] In [47] aneigensystem based method was proposed cooperated withMultiple Layer Perceptron to classify the epileptic seizureshealthy and the seizure-free This approach provided anaverage accuracy of 975 The EMD methods for epilepsydetection achieved 9775 accuracy [6]The Kraskov entropyis also combinedwith the SVMclassifier and provided 9775[22] An automated diagnostics system based on a set ofentropies and fuzzy Sugeno classifier (FSC) achieved accu-racy up to 98 [19] The work presented by [48] developeda method for the epilepsy detection using the EMD Thegenerated IMFs using the EMD were represented as a set ofamplitude and frequency modulated (AMndashFM) signals Thetwo bandwidths namely amplitude modulation bandwidthand frequency modulation bandwidth calculated from theanalytic IMFs have been fed to LS-SVM for classifying
International Journal of Biomedical Imaging 9
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(a) EEG Sample for patient with seizure-free
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(b) EEG Sample for patient with seizures
Figure 3 TQWT decomposition of seizure-free and seizure EEG signals with 119876 = 1 119903 = 3 119895 = 3
1 2 3 4 5 10LEVELS OF DECMPOSISTION (J)
ACCPrecision
SPECRecall
000
020
040
060
080
100
PERC
ENTA
GE
Figure 4 Performance measures of the proposed approach withvaried levels of decomposition
08082084086088
09092094096098
1
1 2 3 5 7 9 10
PERC
ENTA
GE
Q
ACCPrecision
SPECRecall
Figure 5 Performance measures of the proposed approach withvaried levels of decomposition
seizure and nonseizure EEG signals This method achieved9818 average accuracy The LBP-based methods have beencombined with Gabor filter for texture extraction from theEEG signal A k-nearest neighbor classifier was applied andobtained an accuracy of 983 [14] The proposed methodconfirmed its superiority in the total accuracy compared tothe other systems The results which prove the superiority ofthe proposed method compared to the other existed systemsis demonstrated in Figure 8
ACCPrecision
SPECRecall
09091092093094095096097098099
1
1 2 3 4 5PE
RCEN
TAG
ENUMBER OF FEATURES
Figure 6 Performance measures of the original and compactfeatures set
STD
lt= 2873 gt 2873
0
0
ApEn
lt= 016 gt 016
Katz 1
1
lt= 146 gt 146
Figure 7 The classification rules obtained from the proposedsystem to classify the seizure-free (0) and the epileptic seizure (1)signals
Once the system detects the preictal phase the clinicreceives a notification about that patientThemain advantageof the hybrid system from the clinical point of view could besummarized as follows
(i) Classification and detecting of the epileptic seizuresand seizure-free signals from the EEG signal automat-ically
10 International Journal of Biomedical Imaging
90
931195 957 965
975 9775 9775 98 9813 983 99
Baja
j et a
l 20
13
Faus
t et a
l 20
10
Acha
rya e
t al
2009
Mar
tis et
al 2
013
Yuan
et al
201
1
Nag
hsh-
Nilc
hi et
al 2
010
Pach
ori e
t al
2014
Patid
ar et
al 2
017
Acha
rya e
t al
2012
Baja
j et a
l 20
12
Kum
ar et
al 2
015
e p
ropo
sed
met
hod
PROPOSED SYSTESMS USED TO FOR THE COMPARSION
8486889092949698
100
ACCU
RACY
IN P
ERCE
NTA
GE
()
Figure 8 The total accuracy () of various methods employed to detect the disorder of the epilepsy
(ii) The detection of the preictal phase from studying thehealthy and ictal phase of each patient
(iii) The detection of the preictal phase that provided theability to send warning alert to the physicians toprepare the medical assessment for the patient
(iv) The proposed method being robust and reliable as itsperformance was benchmarked using 10-fold cross-validation
(v) The dynamic behavior obtained from the fireflyalgorithm which made the system adaptive to manyfeatures according to each case study
(vi) A few sets of parameters required to analyze the EEGsignal typical three variables (119876 119903 119869)
The limitations of this research were summarized as follows
(i) The limited number of the studied subjects (typically100 per class)
(ii) The diagnosis process that may be reduced because ofdepending on some additional software prerequisites
5 Conclusion
In this paper an automated intelligent CAD tool has beenproposed to classify and detect epileptic seizures and seizure-free EEG signals This method provided an EEG signalanalysis using a hybrid data fusion method The data fusionmethod combined the collected features from two differentperspectives In the first perspective the EEG signal wasconsidered as an image Then the image was converted toa gray image and the GLCM was obtained to extract thetextures of the image such as contrast correlation power andhomogeneity In the second perspective the EEG signal wasdivided into smaller segments using the TQWT to extractthe time and frequency features A set of statistical nonlinear(chaotic) and power spectrum features were obtained from
each segment After the dataset was constructed a featurereduction algorithmbased on firefly optimizationwas used toreduce the irrelevant features and remove redundancy Thenan RFwas trained to classify and predict the epileptic seizuresand seizure-free EEG signals from the dataset The experi-mental results showed that the proposed method achieveda satisfactory degree of 99 accuracy 97 precision 97specificity 98 recall 98 F-measure and 95 MCC at thethird level of decomposition
Data Availability
The dataset used to support the findings of this research wasincluded in the article and can be found as a citation to ldquoRGAndrzejak K Lehnertz F Mormann C Rieke P DavidCE Elger Indications ofNonlinearDeterministic and Finite-Dimensional Structures in Time Series of Brain ElectricalActivity Dependence on Recording Region and Brain StatePhysical Review E 64 (2001) 061907rdquo
Conflicts of Interest
The authors declare that they have no conflicts of interest
References
[1] Epilepsy Fact Sheet 2018 httpwwwwhointmediacentrefactsheetsfs999en
[2] U R Acharya S V Sree G Swapna R J Martis and J S SurildquoAutomated EEG analysis of epilepsy a reviewrdquo Knowledge-Based Systems vol 45 pp 147ndash165 2013
[3] A Puce and M S Hamalainen ldquoA review of issues relatedto data acquisition and analysis in EEGMEG studiesrdquo BrainSciences vol 7 no 6 2017
[4] V Joshi R B Pachori and A Vijesh ldquoClassification of ictaland seizure-free EEG signals using fractional linear predictionrdquoBiomedical Signal Processing and Control vol 9 pp 1ndash5 2014
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
6 International Journal of Biomedical Imaging
(1) InputThe features matrix 119865119898(2) OutputThe optimal solution 119892119887119890119904119905(3) Initialize the firefly swam using a logistic chaotic map(4) Evaluate each firefly 119892119894 using the fitness function f (x)(5) Select the best firefly 119892119887119890119904119905 and the worst one 119892119908119900119903119904119905(6) while termination condition are not reached do(7) Declare an alternative leader firefly as 119878119887119890119904119905 with a competitive fitness and located in different region(8) Obtain an offspring solution 119892119887119890119904119905 using 119878119860(9) for all (firefly 119894 and 119891119894 = 119892119908119900119904119903119905) do(10) for all (firefly 119895 and 119891119895 = 119892119908119900119904119903119905) do(11) if (119868119895 gt 119868119894) then(12) Enhance firefly 119895 using Equation (25) to obtain better offspring candidate firefly denoted 1199091015840119895(13) Replace firefly 119895 with the offspring 1199091015840119895(14) Move the firefly 119895 towards the neighboring and global optimal solutions using Equation (26)(15) end if(16) end for(17) end for(18) Update the worst solution 119892119908119900119903119904119905(19) if 119891(1198921015840119887119890119904119905) gt 119891(119892119887119890119904119905) then(20) 119892119887119890119904119905 larr997888 1198921015840119887119890119904119905(21) end if(22) Rank all fireflies and update the best 119892119887119890119904119905 and worst 119892119908119900119903119904119905 solutions(23) end while(24) return 119892119887119890119904119905Algorithm 1 The pseudocode of the proposed feature reduction algorithm based on firefly optimization and SA
to produce efficient offspring candidates andmemory aware-ness of the best and worst solutions to improve the searchdiversity and prevent local optima The firefly population israndomly initialized using a chaotic logistic map to ensurethe diversely of the candidates and the randomness of eachfirefly Afterwards the fitness function is computed for eachcandidate and identifies both the best and worst solutions as119892119887119890119904119905 and 119892119908119900119904119903119905 respectively For each iteration an alternativecandidate is declared as 119878119887119890119904119905 with a competitive fitnessand located in a different region Both of the best and thealternative candidate sets are used to lead weak solutions toreach the optimal region and prevent local optima problemThe mean of the leader firefly and the alternative one isenhanced using SA algorithm to obtain a better solution1198921015840119887119890119904119905 The improved local and global solutions are used toguide the low lightness fireflies to move towards that strongerlightness The algorithm is repeated until the maximumnumber of iterations is reached or a termination criterion isachieved The behavior of the proposed algorithm is deter-mined by some properties namely the objective function theattractiveness movement step and population diversity Theproposed algorithm is demonstrated in Algorithm 1
TheObjective FunctionThis function is used to evaluate eachcandidate in the algorithm and defined as follows
119891 (119909) = 1199081 times 119886119888119888119906119903119886119888119910 (119909) + 1199082number of features (25)
where 1199081 and 1199082 represent the weights of the classifi-cation accuracy and the number of the selected features
respectively The values of 1199081 and 1199082 are set to 09 and 01respectively as a recommended by [38]
The Attractiveness Movement Step The proposed algorithmused a chaotic logistic map to initialize the firefly populationand hence increases the diversity and avoid local optimaAfter obtaining the global best solution 119892119887119890119904119905 an alternativeleader firefly 119878119887119890119904119905 is declared with a competitive fitness butlocated in a different region Since both leaders are morelikely to discover distinctive search regions this strategyreduces the probability of being trapped in the local optimaIn addition the optimal offspring of themean positions of thetwo leaders and the neighboring brighter candidates are usedto lead the search process and guide the solutions with lowerlight intensity to move towards the optimal region
119909119894 = 119909119894 + 1205730119862119896 (1199091015840119895 minus 119909119894) + 119862119896120576 (1198921015840119887119890119904119905 minus 119909119894) + 1205721015840times sign [119903119886119899119889 minus 05] (26)
1199091015840119895 = 119909119895 + 1205901 (27)
1198921015840119887119890119904119905 = 119898119890119886119899 (119892119887119890119904119905 + 119878119887119890119904119905) + 1205902 (28)
where 1199091015840119895 denotes the offspring candidate with a brighterneighboring solution119909119895 is defined by the SA as shown in (26)and 1198921015840119887119890119904119905 represents the fitter offspring solutions of the meanof the leader firefly and the alternative one as formulated in(27) It worth mentioning that the values of 1205901 and 1205902 aretwo random variables set using the Gaussian distributionThe movement step of the firefly is determined as shown in
International Journal of Biomedical Imaging 7
(1) Input 119894119905119890119903119886119905119894119900119899119904119898119886119909 119892119898119890119886119899 119879119898119886119909(2) OutputThe optimal offspring 119878119887119890119904119905(3) 119878119888 = 119862119903119890119886119905119890119868119899119894119905119878119900119906119897119905119894119900119899119904(119892119898119890119886119899)(4) 119878119887119890119904119905 = 119878119888(5) while (119894 le 119894119905119890119903119886119905119894119900119899119904 119898119886119909) do(6) 119878119894 = 119862119903119890119886119905119890119873119890119894119892ℎ119887119900119903119878119900119897119906119905119894119900119899(119878119888)(7) 119879119888 = 119862119886119897119888119906119897119886119905119890119879119890119898119901119890119903119886119905119906119903119890(119894 119879119898119886119909)(8) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119888)) then(9) 119878119888 = 119878119894(10) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119887119890119904119905)) then(11) 119878119887119890119904119905 = 119878119894(12) end if(13) else if (exp((119888119900119904119905(119878119887119890119904119905) minus 119888119900119904119905(119878119894))119879119888) ge 120590) then(14) 119878119888 = 119878119894(15) end if(16) 119894 = 119894 + 1(17) end while(18) return 119878119887119890119904119905Algorithm 2 The pseudocode of the generation of offspring solutions using SA
(28) where 119862119870 represents the chaotic map variable in themovement step and 120576 denotes the randomized vector definedin the traditional firefly algorithm Parameter 1205721015840 denotes anadaptive step initialized to 05 to control the diversity of thesearch process
Offspring Generation Using SAThe proposed algorithm usedSA for generating better candidates to enhance the searchprocess as much as possible as shown in Algorithm 2 TheSA accepts both of the best solution 119892119887119890119904119905 and the alternativesolution 119878119887119890119904119905 as main inputs then the traditional SA isapplied The better solution generated is accepted by defaultaccording to the SA heuristics On the other hand theweaker solution should be accepted with specific probabilityas shown in (29) where Δ119891 denotes the difference of thefitness (energy) between to candidates and 119879119888 denotes thecurrent temperature A simple linear cooling mechanism isused to control the value of the temperature
119901119909119895 = exp(Δ119891119879119888 ) (29)
Population Diversity In each iteration the worst solution isdetected after the ranking process as shown in Algorithm 1The remaining solutions are guided by the average positionobtained from (30) where 120590 denotes a random value obtainedby Gaussian map
119909119908119900119903119904119905119895 = 119892119887119890119904119905 + 1198781198871198901199041199052 + 120590 (119909119908119900119903119904119905119895 119892119887119890119904119905 + 1198781198871198901199041199052 ) (30)
34 Classification and Learning The random forest (RF) isa successful ensemble approach used in supervised machinelearning to solve classification or regression problems [39]It consists of a collection of decision trees that could actas a single classifier with multiple classification methodsor a method that has several variables Several subsets ofthe training data are supplied to each tree to achieve the
most stable tree classification that results in a generalizedexperience of the classifier The original dataset is dividedinto two parts The first part is used to train each tree bybootstrapping technique The other part is used to evaluatethe accuracy of the classification Each tree is allowed to reachthe maximum depth without tree pruning to obtain a highvariance classifier The splitting process remains until onlyone instance of a single class is dropped from any leaf nodeor a predefined termination condition is achieved Whenthe forest is established the number of subsets remained asa constant The obtained route of traversal from the rootnode to the leaf node is applied to the new instances orthe unlabeled instance for classification The final decisionfor classifying a new instance is provided by determiningeach class that has the most votes from every decision treeThe random forest performs slightly better when comparedwith other classifiers such as discriminant analysis SVM andartificial neural network (ANN) [15 39]
35 Performance Evaluation Various performance formulaswere used to evaluate the effectiveness of a classifierThe sen-sitivity (SEN) or recall specificity (SPEC) accuracy (ACC)F-measure Matthewrsquos correlation coefficient (MCC) andreceiver operating characteristics (ROC) are used to evaluatethe efficiency of the random forest classifier [40ndash42] Theseparameters are defined as follows
119878119864119873 = 119879119875119879119875 + 119865119873 times 100 (31)
119878119875119864119862 = 119879119873119879119873 + 119865119875 times 100 (32)
119860119862119862 = 119879119875 + 119879119873119879119875 + 119879119873 + 119865119875 + 119865119873 times 100 (33)
119865 minus 119898119890119886119904119906119903119890 = 21198791198752119879119875 + 119865119875 + 119865119873 times 100 (34)
8 International Journal of Biomedical Imaging
119872119862119862= (119879119875 times 119879119873) minus (119865119875 times 119865119873)radic(119879119875 + 119865119873) (119879119875 + 119865119875) (119879119873 + 119865119873) (119879119873 + 119865119875)times 100
(35)
119879119875 and 119879119873 represent the total number of an epileptic seizureand seizure-free signals classified correctly respectively Sim-ilarly 119865119875 and 119865119873 represent the total number of epilepticseizures and seizure-free signals classified incorrectly respec-tively Cross-validation has also been used to ensure theclassifier reliability and effectiveness The original dataset issplit into 119896 folds (subsets) for both training and testing Inthis strategy 119896 minus 1 folds were selected randomly to train theclassifier and the remaining folds are used to testing Theoverall performance is calculated as the average of each foldIn this work the experiment is repeated 10 times with tenfoldcross-validation
4 Results and Discussion
The benchmark dataset used in this investigation wasacquired by the University of Bonn [28]The dataset containsthree different categories ie preictal healthy and ictalrecorded using a single channel for a 236 s duration Bothnormal and preictal conditions were collected from 200 casestudies and 100 for the ictal state The normal condition isacquired from five healthy volunteers using the international10ndash20 system standard with each volunteer in a relaxed-awake state with eyes open and closed (100 cases per each set)[28] The ictal data were collected from five patients duringtheir epileptic seizures The preictal represents the EEG datacollected from the same five patients with no seizures It isworth mentioning that all EEG signals were acquired using a128 channel amplifier with sampling rate equal 17361Hz [28]Finally a bandpass filter with 05340Hz sim12 dBoctave wasapplied as a filter
The proposed approach for automatic detection of epilep-tic seizures and seizure-free patients was implemented usingMATLAB software A TQWT comparison between theseizure-free and the epileptic seizures patients are shown inFigure 3 with 119876 = 1 119903 = 3 and 119869 = 3 It can be observedthat both of the amplitudes and the frequency of the epilepticseizure are much higher than the healthy one Moreover theoscillatory behavior of the epileptic patient is higher than thehealthy one The value of the parameter 119903 is set to three toprevent any excessive ringing of the wavelet as suggested by[22] MATLAB for the TQWT toolbox is available for publicaccess at httpeewebpolyeduiselesniTQWT
After the construction of the dataset the feature reduc-tion was applied using the firefly algorithm to remove theredundant and irrelevant features The number of the datasegments was determined by the parameters 119876 119903 and 119869 Bytuning these parameters the number of the data segmentswas varied and thus the training process was adopted Thetrial and error approach was used to set the value of theseparameters The experiment was implemented on variousvalues of J to obtain the best level of decomposition The
performance measures were calculated for each level ofdecomposition As shown in Figure 4 the best value of thevariable 119869 was from two to three Moreover the best value ofthe parameter119876was found to be one as shown in Figure 5 Allthe variables remained constant while changing the Q-factorto obtain the best value
Then the firefly algorithm with a population size of 20fireflies mutation probability of 001 and light absorptionequal to 01 was applied to reduce the feature setThe result ofthe compact set of features is obtained after 100 iterations andshown in Figure 6 are the number of features obtained by thefirefly algorithm and their corresponding accuracy precisionspecificity and the recall In the last step the feature set isfed to a random forest classifier to obtain the seizure-free andseizure conditionsThe random forest classifier obtained 98accuracy 97 precision 97 specificity 98 recall 98 F-measure and 95 MCC at the third level of decompositionand the Q-factor equals 1
The firefly algorithm reduced the search space into threefeatures These features could replace the original datasetwhich minimizes the processing time The compact set offeature contains the 119878119879119863 119860119901119864119899 and the 119870119860119879119885 the classi-fication rules are based on these features The classificationrules obtained by the proposed system are shown in Figure 7where the decision tree consists of 4 leaves and 7 nodes andthe final decision represents the seizure-free and the epilepticseizure denoted by 01 respectively
A comparative study of the proposed hybrid epilepsydetection approach and other existing classification systemshas been performed in terms of the total accuracy A novelmethod based on the EMDs was proposed to detect theepileptic seizures of epilepsy This method used the Hilberttransformation of IMFs obtained by EMD process thatprovided an analytic signal representation of IMFs [12] Theclassification rules obtained by this method achieved anaccuracy of 90 The usage of frequency domain featuresand Burgrsquos method obtained 9311 accuracy with SVMclassifier [43] Nonlinear features have been used with aGaussianmixturemodel classifier and achieved 95accuracy[44] A decision tree classifier is used with energy fractaldimension and sample entropy and provided 957 [45]A combination of ApEn and the Hurst exponent has beenused to detect the diagnostics of epilepsy and produced 965accuracy with SVM classifier and ANN [46] In [47] aneigensystem based method was proposed cooperated withMultiple Layer Perceptron to classify the epileptic seizureshealthy and the seizure-free This approach provided anaverage accuracy of 975 The EMD methods for epilepsydetection achieved 9775 accuracy [6]The Kraskov entropyis also combinedwith the SVMclassifier and provided 9775[22] An automated diagnostics system based on a set ofentropies and fuzzy Sugeno classifier (FSC) achieved accu-racy up to 98 [19] The work presented by [48] developeda method for the epilepsy detection using the EMD Thegenerated IMFs using the EMD were represented as a set ofamplitude and frequency modulated (AMndashFM) signals Thetwo bandwidths namely amplitude modulation bandwidthand frequency modulation bandwidth calculated from theanalytic IMFs have been fed to LS-SVM for classifying
International Journal of Biomedical Imaging 9
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(a) EEG Sample for patient with seizure-free
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(b) EEG Sample for patient with seizures
Figure 3 TQWT decomposition of seizure-free and seizure EEG signals with 119876 = 1 119903 = 3 119895 = 3
1 2 3 4 5 10LEVELS OF DECMPOSISTION (J)
ACCPrecision
SPECRecall
000
020
040
060
080
100
PERC
ENTA
GE
Figure 4 Performance measures of the proposed approach withvaried levels of decomposition
08082084086088
09092094096098
1
1 2 3 5 7 9 10
PERC
ENTA
GE
Q
ACCPrecision
SPECRecall
Figure 5 Performance measures of the proposed approach withvaried levels of decomposition
seizure and nonseizure EEG signals This method achieved9818 average accuracy The LBP-based methods have beencombined with Gabor filter for texture extraction from theEEG signal A k-nearest neighbor classifier was applied andobtained an accuracy of 983 [14] The proposed methodconfirmed its superiority in the total accuracy compared tothe other systems The results which prove the superiority ofthe proposed method compared to the other existed systemsis demonstrated in Figure 8
ACCPrecision
SPECRecall
09091092093094095096097098099
1
1 2 3 4 5PE
RCEN
TAG
ENUMBER OF FEATURES
Figure 6 Performance measures of the original and compactfeatures set
STD
lt= 2873 gt 2873
0
0
ApEn
lt= 016 gt 016
Katz 1
1
lt= 146 gt 146
Figure 7 The classification rules obtained from the proposedsystem to classify the seizure-free (0) and the epileptic seizure (1)signals
Once the system detects the preictal phase the clinicreceives a notification about that patientThemain advantageof the hybrid system from the clinical point of view could besummarized as follows
(i) Classification and detecting of the epileptic seizuresand seizure-free signals from the EEG signal automat-ically
10 International Journal of Biomedical Imaging
90
931195 957 965
975 9775 9775 98 9813 983 99
Baja
j et a
l 20
13
Faus
t et a
l 20
10
Acha
rya e
t al
2009
Mar
tis et
al 2
013
Yuan
et al
201
1
Nag
hsh-
Nilc
hi et
al 2
010
Pach
ori e
t al
2014
Patid
ar et
al 2
017
Acha
rya e
t al
2012
Baja
j et a
l 20
12
Kum
ar et
al 2
015
e p
ropo
sed
met
hod
PROPOSED SYSTESMS USED TO FOR THE COMPARSION
8486889092949698
100
ACCU
RACY
IN P
ERCE
NTA
GE
()
Figure 8 The total accuracy () of various methods employed to detect the disorder of the epilepsy
(ii) The detection of the preictal phase from studying thehealthy and ictal phase of each patient
(iii) The detection of the preictal phase that provided theability to send warning alert to the physicians toprepare the medical assessment for the patient
(iv) The proposed method being robust and reliable as itsperformance was benchmarked using 10-fold cross-validation
(v) The dynamic behavior obtained from the fireflyalgorithm which made the system adaptive to manyfeatures according to each case study
(vi) A few sets of parameters required to analyze the EEGsignal typical three variables (119876 119903 119869)
The limitations of this research were summarized as follows
(i) The limited number of the studied subjects (typically100 per class)
(ii) The diagnosis process that may be reduced because ofdepending on some additional software prerequisites
5 Conclusion
In this paper an automated intelligent CAD tool has beenproposed to classify and detect epileptic seizures and seizure-free EEG signals This method provided an EEG signalanalysis using a hybrid data fusion method The data fusionmethod combined the collected features from two differentperspectives In the first perspective the EEG signal wasconsidered as an image Then the image was converted toa gray image and the GLCM was obtained to extract thetextures of the image such as contrast correlation power andhomogeneity In the second perspective the EEG signal wasdivided into smaller segments using the TQWT to extractthe time and frequency features A set of statistical nonlinear(chaotic) and power spectrum features were obtained from
each segment After the dataset was constructed a featurereduction algorithmbased on firefly optimizationwas used toreduce the irrelevant features and remove redundancy Thenan RFwas trained to classify and predict the epileptic seizuresand seizure-free EEG signals from the dataset The experi-mental results showed that the proposed method achieveda satisfactory degree of 99 accuracy 97 precision 97specificity 98 recall 98 F-measure and 95 MCC at thethird level of decomposition
Data Availability
The dataset used to support the findings of this research wasincluded in the article and can be found as a citation to ldquoRGAndrzejak K Lehnertz F Mormann C Rieke P DavidCE Elger Indications ofNonlinearDeterministic and Finite-Dimensional Structures in Time Series of Brain ElectricalActivity Dependence on Recording Region and Brain StatePhysical Review E 64 (2001) 061907rdquo
Conflicts of Interest
The authors declare that they have no conflicts of interest
References
[1] Epilepsy Fact Sheet 2018 httpwwwwhointmediacentrefactsheetsfs999en
[2] U R Acharya S V Sree G Swapna R J Martis and J S SurildquoAutomated EEG analysis of epilepsy a reviewrdquo Knowledge-Based Systems vol 45 pp 147ndash165 2013
[3] A Puce and M S Hamalainen ldquoA review of issues relatedto data acquisition and analysis in EEGMEG studiesrdquo BrainSciences vol 7 no 6 2017
[4] V Joshi R B Pachori and A Vijesh ldquoClassification of ictaland seizure-free EEG signals using fractional linear predictionrdquoBiomedical Signal Processing and Control vol 9 pp 1ndash5 2014
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
International Journal of Biomedical Imaging 7
(1) Input 119894119905119890119903119886119905119894119900119899119904119898119886119909 119892119898119890119886119899 119879119898119886119909(2) OutputThe optimal offspring 119878119887119890119904119905(3) 119878119888 = 119862119903119890119886119905119890119868119899119894119905119878119900119906119897119905119894119900119899119904(119892119898119890119886119899)(4) 119878119887119890119904119905 = 119878119888(5) while (119894 le 119894119905119890119903119886119905119894119900119899119904 119898119886119909) do(6) 119878119894 = 119862119903119890119886119905119890119873119890119894119892ℎ119887119900119903119878119900119897119906119905119894119900119899(119878119888)(7) 119879119888 = 119862119886119897119888119906119897119886119905119890119879119890119898119901119890119903119886119905119906119903119890(119894 119879119898119886119909)(8) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119888)) then(9) 119878119888 = 119878119894(10) if (119888119900119904119905(119878119894) le 119888119900119904119905(119878119887119890119904119905)) then(11) 119878119887119890119904119905 = 119878119894(12) end if(13) else if (exp((119888119900119904119905(119878119887119890119904119905) minus 119888119900119904119905(119878119894))119879119888) ge 120590) then(14) 119878119888 = 119878119894(15) end if(16) 119894 = 119894 + 1(17) end while(18) return 119878119887119890119904119905Algorithm 2 The pseudocode of the generation of offspring solutions using SA
(28) where 119862119870 represents the chaotic map variable in themovement step and 120576 denotes the randomized vector definedin the traditional firefly algorithm Parameter 1205721015840 denotes anadaptive step initialized to 05 to control the diversity of thesearch process
Offspring Generation Using SAThe proposed algorithm usedSA for generating better candidates to enhance the searchprocess as much as possible as shown in Algorithm 2 TheSA accepts both of the best solution 119892119887119890119904119905 and the alternativesolution 119878119887119890119904119905 as main inputs then the traditional SA isapplied The better solution generated is accepted by defaultaccording to the SA heuristics On the other hand theweaker solution should be accepted with specific probabilityas shown in (29) where Δ119891 denotes the difference of thefitness (energy) between to candidates and 119879119888 denotes thecurrent temperature A simple linear cooling mechanism isused to control the value of the temperature
119901119909119895 = exp(Δ119891119879119888 ) (29)
Population Diversity In each iteration the worst solution isdetected after the ranking process as shown in Algorithm 1The remaining solutions are guided by the average positionobtained from (30) where 120590 denotes a random value obtainedby Gaussian map
119909119908119900119903119904119905119895 = 119892119887119890119904119905 + 1198781198871198901199041199052 + 120590 (119909119908119900119903119904119905119895 119892119887119890119904119905 + 1198781198871198901199041199052 ) (30)
34 Classification and Learning The random forest (RF) isa successful ensemble approach used in supervised machinelearning to solve classification or regression problems [39]It consists of a collection of decision trees that could actas a single classifier with multiple classification methodsor a method that has several variables Several subsets ofthe training data are supplied to each tree to achieve the
most stable tree classification that results in a generalizedexperience of the classifier The original dataset is dividedinto two parts The first part is used to train each tree bybootstrapping technique The other part is used to evaluatethe accuracy of the classification Each tree is allowed to reachthe maximum depth without tree pruning to obtain a highvariance classifier The splitting process remains until onlyone instance of a single class is dropped from any leaf nodeor a predefined termination condition is achieved Whenthe forest is established the number of subsets remained asa constant The obtained route of traversal from the rootnode to the leaf node is applied to the new instances orthe unlabeled instance for classification The final decisionfor classifying a new instance is provided by determiningeach class that has the most votes from every decision treeThe random forest performs slightly better when comparedwith other classifiers such as discriminant analysis SVM andartificial neural network (ANN) [15 39]
35 Performance Evaluation Various performance formulaswere used to evaluate the effectiveness of a classifierThe sen-sitivity (SEN) or recall specificity (SPEC) accuracy (ACC)F-measure Matthewrsquos correlation coefficient (MCC) andreceiver operating characteristics (ROC) are used to evaluatethe efficiency of the random forest classifier [40ndash42] Theseparameters are defined as follows
119878119864119873 = 119879119875119879119875 + 119865119873 times 100 (31)
119878119875119864119862 = 119879119873119879119873 + 119865119875 times 100 (32)
119860119862119862 = 119879119875 + 119879119873119879119875 + 119879119873 + 119865119875 + 119865119873 times 100 (33)
119865 minus 119898119890119886119904119906119903119890 = 21198791198752119879119875 + 119865119875 + 119865119873 times 100 (34)
8 International Journal of Biomedical Imaging
119872119862119862= (119879119875 times 119879119873) minus (119865119875 times 119865119873)radic(119879119875 + 119865119873) (119879119875 + 119865119875) (119879119873 + 119865119873) (119879119873 + 119865119875)times 100
(35)
119879119875 and 119879119873 represent the total number of an epileptic seizureand seizure-free signals classified correctly respectively Sim-ilarly 119865119875 and 119865119873 represent the total number of epilepticseizures and seizure-free signals classified incorrectly respec-tively Cross-validation has also been used to ensure theclassifier reliability and effectiveness The original dataset issplit into 119896 folds (subsets) for both training and testing Inthis strategy 119896 minus 1 folds were selected randomly to train theclassifier and the remaining folds are used to testing Theoverall performance is calculated as the average of each foldIn this work the experiment is repeated 10 times with tenfoldcross-validation
4 Results and Discussion
The benchmark dataset used in this investigation wasacquired by the University of Bonn [28]The dataset containsthree different categories ie preictal healthy and ictalrecorded using a single channel for a 236 s duration Bothnormal and preictal conditions were collected from 200 casestudies and 100 for the ictal state The normal condition isacquired from five healthy volunteers using the international10ndash20 system standard with each volunteer in a relaxed-awake state with eyes open and closed (100 cases per each set)[28] The ictal data were collected from five patients duringtheir epileptic seizures The preictal represents the EEG datacollected from the same five patients with no seizures It isworth mentioning that all EEG signals were acquired using a128 channel amplifier with sampling rate equal 17361Hz [28]Finally a bandpass filter with 05340Hz sim12 dBoctave wasapplied as a filter
The proposed approach for automatic detection of epilep-tic seizures and seizure-free patients was implemented usingMATLAB software A TQWT comparison between theseizure-free and the epileptic seizures patients are shown inFigure 3 with 119876 = 1 119903 = 3 and 119869 = 3 It can be observedthat both of the amplitudes and the frequency of the epilepticseizure are much higher than the healthy one Moreover theoscillatory behavior of the epileptic patient is higher than thehealthy one The value of the parameter 119903 is set to three toprevent any excessive ringing of the wavelet as suggested by[22] MATLAB for the TQWT toolbox is available for publicaccess at httpeewebpolyeduiselesniTQWT
After the construction of the dataset the feature reduc-tion was applied using the firefly algorithm to remove theredundant and irrelevant features The number of the datasegments was determined by the parameters 119876 119903 and 119869 Bytuning these parameters the number of the data segmentswas varied and thus the training process was adopted Thetrial and error approach was used to set the value of theseparameters The experiment was implemented on variousvalues of J to obtain the best level of decomposition The
performance measures were calculated for each level ofdecomposition As shown in Figure 4 the best value of thevariable 119869 was from two to three Moreover the best value ofthe parameter119876was found to be one as shown in Figure 5 Allthe variables remained constant while changing the Q-factorto obtain the best value
Then the firefly algorithm with a population size of 20fireflies mutation probability of 001 and light absorptionequal to 01 was applied to reduce the feature setThe result ofthe compact set of features is obtained after 100 iterations andshown in Figure 6 are the number of features obtained by thefirefly algorithm and their corresponding accuracy precisionspecificity and the recall In the last step the feature set isfed to a random forest classifier to obtain the seizure-free andseizure conditionsThe random forest classifier obtained 98accuracy 97 precision 97 specificity 98 recall 98 F-measure and 95 MCC at the third level of decompositionand the Q-factor equals 1
The firefly algorithm reduced the search space into threefeatures These features could replace the original datasetwhich minimizes the processing time The compact set offeature contains the 119878119879119863 119860119901119864119899 and the 119870119860119879119885 the classi-fication rules are based on these features The classificationrules obtained by the proposed system are shown in Figure 7where the decision tree consists of 4 leaves and 7 nodes andthe final decision represents the seizure-free and the epilepticseizure denoted by 01 respectively
A comparative study of the proposed hybrid epilepsydetection approach and other existing classification systemshas been performed in terms of the total accuracy A novelmethod based on the EMDs was proposed to detect theepileptic seizures of epilepsy This method used the Hilberttransformation of IMFs obtained by EMD process thatprovided an analytic signal representation of IMFs [12] Theclassification rules obtained by this method achieved anaccuracy of 90 The usage of frequency domain featuresand Burgrsquos method obtained 9311 accuracy with SVMclassifier [43] Nonlinear features have been used with aGaussianmixturemodel classifier and achieved 95accuracy[44] A decision tree classifier is used with energy fractaldimension and sample entropy and provided 957 [45]A combination of ApEn and the Hurst exponent has beenused to detect the diagnostics of epilepsy and produced 965accuracy with SVM classifier and ANN [46] In [47] aneigensystem based method was proposed cooperated withMultiple Layer Perceptron to classify the epileptic seizureshealthy and the seizure-free This approach provided anaverage accuracy of 975 The EMD methods for epilepsydetection achieved 9775 accuracy [6]The Kraskov entropyis also combinedwith the SVMclassifier and provided 9775[22] An automated diagnostics system based on a set ofentropies and fuzzy Sugeno classifier (FSC) achieved accu-racy up to 98 [19] The work presented by [48] developeda method for the epilepsy detection using the EMD Thegenerated IMFs using the EMD were represented as a set ofamplitude and frequency modulated (AMndashFM) signals Thetwo bandwidths namely amplitude modulation bandwidthand frequency modulation bandwidth calculated from theanalytic IMFs have been fed to LS-SVM for classifying
International Journal of Biomedical Imaging 9
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(a) EEG Sample for patient with seizure-free
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(b) EEG Sample for patient with seizures
Figure 3 TQWT decomposition of seizure-free and seizure EEG signals with 119876 = 1 119903 = 3 119895 = 3
1 2 3 4 5 10LEVELS OF DECMPOSISTION (J)
ACCPrecision
SPECRecall
000
020
040
060
080
100
PERC
ENTA
GE
Figure 4 Performance measures of the proposed approach withvaried levels of decomposition
08082084086088
09092094096098
1
1 2 3 5 7 9 10
PERC
ENTA
GE
Q
ACCPrecision
SPECRecall
Figure 5 Performance measures of the proposed approach withvaried levels of decomposition
seizure and nonseizure EEG signals This method achieved9818 average accuracy The LBP-based methods have beencombined with Gabor filter for texture extraction from theEEG signal A k-nearest neighbor classifier was applied andobtained an accuracy of 983 [14] The proposed methodconfirmed its superiority in the total accuracy compared tothe other systems The results which prove the superiority ofthe proposed method compared to the other existed systemsis demonstrated in Figure 8
ACCPrecision
SPECRecall
09091092093094095096097098099
1
1 2 3 4 5PE
RCEN
TAG
ENUMBER OF FEATURES
Figure 6 Performance measures of the original and compactfeatures set
STD
lt= 2873 gt 2873
0
0
ApEn
lt= 016 gt 016
Katz 1
1
lt= 146 gt 146
Figure 7 The classification rules obtained from the proposedsystem to classify the seizure-free (0) and the epileptic seizure (1)signals
Once the system detects the preictal phase the clinicreceives a notification about that patientThemain advantageof the hybrid system from the clinical point of view could besummarized as follows
(i) Classification and detecting of the epileptic seizuresand seizure-free signals from the EEG signal automat-ically
10 International Journal of Biomedical Imaging
90
931195 957 965
975 9775 9775 98 9813 983 99
Baja
j et a
l 20
13
Faus
t et a
l 20
10
Acha
rya e
t al
2009
Mar
tis et
al 2
013
Yuan
et al
201
1
Nag
hsh-
Nilc
hi et
al 2
010
Pach
ori e
t al
2014
Patid
ar et
al 2
017
Acha
rya e
t al
2012
Baja
j et a
l 20
12
Kum
ar et
al 2
015
e p
ropo
sed
met
hod
PROPOSED SYSTESMS USED TO FOR THE COMPARSION
8486889092949698
100
ACCU
RACY
IN P
ERCE
NTA
GE
()
Figure 8 The total accuracy () of various methods employed to detect the disorder of the epilepsy
(ii) The detection of the preictal phase from studying thehealthy and ictal phase of each patient
(iii) The detection of the preictal phase that provided theability to send warning alert to the physicians toprepare the medical assessment for the patient
(iv) The proposed method being robust and reliable as itsperformance was benchmarked using 10-fold cross-validation
(v) The dynamic behavior obtained from the fireflyalgorithm which made the system adaptive to manyfeatures according to each case study
(vi) A few sets of parameters required to analyze the EEGsignal typical three variables (119876 119903 119869)
The limitations of this research were summarized as follows
(i) The limited number of the studied subjects (typically100 per class)
(ii) The diagnosis process that may be reduced because ofdepending on some additional software prerequisites
5 Conclusion
In this paper an automated intelligent CAD tool has beenproposed to classify and detect epileptic seizures and seizure-free EEG signals This method provided an EEG signalanalysis using a hybrid data fusion method The data fusionmethod combined the collected features from two differentperspectives In the first perspective the EEG signal wasconsidered as an image Then the image was converted toa gray image and the GLCM was obtained to extract thetextures of the image such as contrast correlation power andhomogeneity In the second perspective the EEG signal wasdivided into smaller segments using the TQWT to extractthe time and frequency features A set of statistical nonlinear(chaotic) and power spectrum features were obtained from
each segment After the dataset was constructed a featurereduction algorithmbased on firefly optimizationwas used toreduce the irrelevant features and remove redundancy Thenan RFwas trained to classify and predict the epileptic seizuresand seizure-free EEG signals from the dataset The experi-mental results showed that the proposed method achieveda satisfactory degree of 99 accuracy 97 precision 97specificity 98 recall 98 F-measure and 95 MCC at thethird level of decomposition
Data Availability
The dataset used to support the findings of this research wasincluded in the article and can be found as a citation to ldquoRGAndrzejak K Lehnertz F Mormann C Rieke P DavidCE Elger Indications ofNonlinearDeterministic and Finite-Dimensional Structures in Time Series of Brain ElectricalActivity Dependence on Recording Region and Brain StatePhysical Review E 64 (2001) 061907rdquo
Conflicts of Interest
The authors declare that they have no conflicts of interest
References
[1] Epilepsy Fact Sheet 2018 httpwwwwhointmediacentrefactsheetsfs999en
[2] U R Acharya S V Sree G Swapna R J Martis and J S SurildquoAutomated EEG analysis of epilepsy a reviewrdquo Knowledge-Based Systems vol 45 pp 147ndash165 2013
[3] A Puce and M S Hamalainen ldquoA review of issues relatedto data acquisition and analysis in EEGMEG studiesrdquo BrainSciences vol 7 no 6 2017
[4] V Joshi R B Pachori and A Vijesh ldquoClassification of ictaland seizure-free EEG signals using fractional linear predictionrdquoBiomedical Signal Processing and Control vol 9 pp 1ndash5 2014
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
8 International Journal of Biomedical Imaging
119872119862119862= (119879119875 times 119879119873) minus (119865119875 times 119865119873)radic(119879119875 + 119865119873) (119879119875 + 119865119875) (119879119873 + 119865119873) (119879119873 + 119865119875)times 100
(35)
119879119875 and 119879119873 represent the total number of an epileptic seizureand seizure-free signals classified correctly respectively Sim-ilarly 119865119875 and 119865119873 represent the total number of epilepticseizures and seizure-free signals classified incorrectly respec-tively Cross-validation has also been used to ensure theclassifier reliability and effectiveness The original dataset issplit into 119896 folds (subsets) for both training and testing Inthis strategy 119896 minus 1 folds were selected randomly to train theclassifier and the remaining folds are used to testing Theoverall performance is calculated as the average of each foldIn this work the experiment is repeated 10 times with tenfoldcross-validation
4 Results and Discussion
The benchmark dataset used in this investigation wasacquired by the University of Bonn [28]The dataset containsthree different categories ie preictal healthy and ictalrecorded using a single channel for a 236 s duration Bothnormal and preictal conditions were collected from 200 casestudies and 100 for the ictal state The normal condition isacquired from five healthy volunteers using the international10ndash20 system standard with each volunteer in a relaxed-awake state with eyes open and closed (100 cases per each set)[28] The ictal data were collected from five patients duringtheir epileptic seizures The preictal represents the EEG datacollected from the same five patients with no seizures It isworth mentioning that all EEG signals were acquired using a128 channel amplifier with sampling rate equal 17361Hz [28]Finally a bandpass filter with 05340Hz sim12 dBoctave wasapplied as a filter
The proposed approach for automatic detection of epilep-tic seizures and seizure-free patients was implemented usingMATLAB software A TQWT comparison between theseizure-free and the epileptic seizures patients are shown inFigure 3 with 119876 = 1 119903 = 3 and 119869 = 3 It can be observedthat both of the amplitudes and the frequency of the epilepticseizure are much higher than the healthy one Moreover theoscillatory behavior of the epileptic patient is higher than thehealthy one The value of the parameter 119903 is set to three toprevent any excessive ringing of the wavelet as suggested by[22] MATLAB for the TQWT toolbox is available for publicaccess at httpeewebpolyeduiselesniTQWT
After the construction of the dataset the feature reduc-tion was applied using the firefly algorithm to remove theredundant and irrelevant features The number of the datasegments was determined by the parameters 119876 119903 and 119869 Bytuning these parameters the number of the data segmentswas varied and thus the training process was adopted Thetrial and error approach was used to set the value of theseparameters The experiment was implemented on variousvalues of J to obtain the best level of decomposition The
performance measures were calculated for each level ofdecomposition As shown in Figure 4 the best value of thevariable 119869 was from two to three Moreover the best value ofthe parameter119876was found to be one as shown in Figure 5 Allthe variables remained constant while changing the Q-factorto obtain the best value
Then the firefly algorithm with a population size of 20fireflies mutation probability of 001 and light absorptionequal to 01 was applied to reduce the feature setThe result ofthe compact set of features is obtained after 100 iterations andshown in Figure 6 are the number of features obtained by thefirefly algorithm and their corresponding accuracy precisionspecificity and the recall In the last step the feature set isfed to a random forest classifier to obtain the seizure-free andseizure conditionsThe random forest classifier obtained 98accuracy 97 precision 97 specificity 98 recall 98 F-measure and 95 MCC at the third level of decompositionand the Q-factor equals 1
The firefly algorithm reduced the search space into threefeatures These features could replace the original datasetwhich minimizes the processing time The compact set offeature contains the 119878119879119863 119860119901119864119899 and the 119870119860119879119885 the classi-fication rules are based on these features The classificationrules obtained by the proposed system are shown in Figure 7where the decision tree consists of 4 leaves and 7 nodes andthe final decision represents the seizure-free and the epilepticseizure denoted by 01 respectively
A comparative study of the proposed hybrid epilepsydetection approach and other existing classification systemshas been performed in terms of the total accuracy A novelmethod based on the EMDs was proposed to detect theepileptic seizures of epilepsy This method used the Hilberttransformation of IMFs obtained by EMD process thatprovided an analytic signal representation of IMFs [12] Theclassification rules obtained by this method achieved anaccuracy of 90 The usage of frequency domain featuresand Burgrsquos method obtained 9311 accuracy with SVMclassifier [43] Nonlinear features have been used with aGaussianmixturemodel classifier and achieved 95accuracy[44] A decision tree classifier is used with energy fractaldimension and sample entropy and provided 957 [45]A combination of ApEn and the Hurst exponent has beenused to detect the diagnostics of epilepsy and produced 965accuracy with SVM classifier and ANN [46] In [47] aneigensystem based method was proposed cooperated withMultiple Layer Perceptron to classify the epileptic seizureshealthy and the seizure-free This approach provided anaverage accuracy of 975 The EMD methods for epilepsydetection achieved 9775 accuracy [6]The Kraskov entropyis also combinedwith the SVMclassifier and provided 9775[22] An automated diagnostics system based on a set ofentropies and fuzzy Sugeno classifier (FSC) achieved accu-racy up to 98 [19] The work presented by [48] developeda method for the epilepsy detection using the EMD Thegenerated IMFs using the EMD were represented as a set ofamplitude and frequency modulated (AMndashFM) signals Thetwo bandwidths namely amplitude modulation bandwidthand frequency modulation bandwidth calculated from theanalytic IMFs have been fed to LS-SVM for classifying
International Journal of Biomedical Imaging 9
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(a) EEG Sample for patient with seizure-free
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(b) EEG Sample for patient with seizures
Figure 3 TQWT decomposition of seizure-free and seizure EEG signals with 119876 = 1 119903 = 3 119895 = 3
1 2 3 4 5 10LEVELS OF DECMPOSISTION (J)
ACCPrecision
SPECRecall
000
020
040
060
080
100
PERC
ENTA
GE
Figure 4 Performance measures of the proposed approach withvaried levels of decomposition
08082084086088
09092094096098
1
1 2 3 5 7 9 10
PERC
ENTA
GE
Q
ACCPrecision
SPECRecall
Figure 5 Performance measures of the proposed approach withvaried levels of decomposition
seizure and nonseizure EEG signals This method achieved9818 average accuracy The LBP-based methods have beencombined with Gabor filter for texture extraction from theEEG signal A k-nearest neighbor classifier was applied andobtained an accuracy of 983 [14] The proposed methodconfirmed its superiority in the total accuracy compared tothe other systems The results which prove the superiority ofthe proposed method compared to the other existed systemsis demonstrated in Figure 8
ACCPrecision
SPECRecall
09091092093094095096097098099
1
1 2 3 4 5PE
RCEN
TAG
ENUMBER OF FEATURES
Figure 6 Performance measures of the original and compactfeatures set
STD
lt= 2873 gt 2873
0
0
ApEn
lt= 016 gt 016
Katz 1
1
lt= 146 gt 146
Figure 7 The classification rules obtained from the proposedsystem to classify the seizure-free (0) and the epileptic seizure (1)signals
Once the system detects the preictal phase the clinicreceives a notification about that patientThemain advantageof the hybrid system from the clinical point of view could besummarized as follows
(i) Classification and detecting of the epileptic seizuresand seizure-free signals from the EEG signal automat-ically
10 International Journal of Biomedical Imaging
90
931195 957 965
975 9775 9775 98 9813 983 99
Baja
j et a
l 20
13
Faus
t et a
l 20
10
Acha
rya e
t al
2009
Mar
tis et
al 2
013
Yuan
et al
201
1
Nag
hsh-
Nilc
hi et
al 2
010
Pach
ori e
t al
2014
Patid
ar et
al 2
017
Acha
rya e
t al
2012
Baja
j et a
l 20
12
Kum
ar et
al 2
015
e p
ropo
sed
met
hod
PROPOSED SYSTESMS USED TO FOR THE COMPARSION
8486889092949698
100
ACCU
RACY
IN P
ERCE
NTA
GE
()
Figure 8 The total accuracy () of various methods employed to detect the disorder of the epilepsy
(ii) The detection of the preictal phase from studying thehealthy and ictal phase of each patient
(iii) The detection of the preictal phase that provided theability to send warning alert to the physicians toprepare the medical assessment for the patient
(iv) The proposed method being robust and reliable as itsperformance was benchmarked using 10-fold cross-validation
(v) The dynamic behavior obtained from the fireflyalgorithm which made the system adaptive to manyfeatures according to each case study
(vi) A few sets of parameters required to analyze the EEGsignal typical three variables (119876 119903 119869)
The limitations of this research were summarized as follows
(i) The limited number of the studied subjects (typically100 per class)
(ii) The diagnosis process that may be reduced because ofdepending on some additional software prerequisites
5 Conclusion
In this paper an automated intelligent CAD tool has beenproposed to classify and detect epileptic seizures and seizure-free EEG signals This method provided an EEG signalanalysis using a hybrid data fusion method The data fusionmethod combined the collected features from two differentperspectives In the first perspective the EEG signal wasconsidered as an image Then the image was converted toa gray image and the GLCM was obtained to extract thetextures of the image such as contrast correlation power andhomogeneity In the second perspective the EEG signal wasdivided into smaller segments using the TQWT to extractthe time and frequency features A set of statistical nonlinear(chaotic) and power spectrum features were obtained from
each segment After the dataset was constructed a featurereduction algorithmbased on firefly optimizationwas used toreduce the irrelevant features and remove redundancy Thenan RFwas trained to classify and predict the epileptic seizuresand seizure-free EEG signals from the dataset The experi-mental results showed that the proposed method achieveda satisfactory degree of 99 accuracy 97 precision 97specificity 98 recall 98 F-measure and 95 MCC at thethird level of decomposition
Data Availability
The dataset used to support the findings of this research wasincluded in the article and can be found as a citation to ldquoRGAndrzejak K Lehnertz F Mormann C Rieke P DavidCE Elger Indications ofNonlinearDeterministic and Finite-Dimensional Structures in Time Series of Brain ElectricalActivity Dependence on Recording Region and Brain StatePhysical Review E 64 (2001) 061907rdquo
Conflicts of Interest
The authors declare that they have no conflicts of interest
References
[1] Epilepsy Fact Sheet 2018 httpwwwwhointmediacentrefactsheetsfs999en
[2] U R Acharya S V Sree G Swapna R J Martis and J S SurildquoAutomated EEG analysis of epilepsy a reviewrdquo Knowledge-Based Systems vol 45 pp 147ndash165 2013
[3] A Puce and M S Hamalainen ldquoA review of issues relatedto data acquisition and analysis in EEGMEG studiesrdquo BrainSciences vol 7 no 6 2017
[4] V Joshi R B Pachori and A Vijesh ldquoClassification of ictaland seizure-free EEG signals using fractional linear predictionrdquoBiomedical Signal Processing and Control vol 9 pp 1ndash5 2014
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
International Journal of Biomedical Imaging 9
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(a) EEG Sample for patient with seizure-free
4321
SUBB
AN
D
500 1000 1500 2000 2500 3000 3500 40000TIME (SAMPLES)
(b) EEG Sample for patient with seizures
Figure 3 TQWT decomposition of seizure-free and seizure EEG signals with 119876 = 1 119903 = 3 119895 = 3
1 2 3 4 5 10LEVELS OF DECMPOSISTION (J)
ACCPrecision
SPECRecall
000
020
040
060
080
100
PERC
ENTA
GE
Figure 4 Performance measures of the proposed approach withvaried levels of decomposition
08082084086088
09092094096098
1
1 2 3 5 7 9 10
PERC
ENTA
GE
Q
ACCPrecision
SPECRecall
Figure 5 Performance measures of the proposed approach withvaried levels of decomposition
seizure and nonseizure EEG signals This method achieved9818 average accuracy The LBP-based methods have beencombined with Gabor filter for texture extraction from theEEG signal A k-nearest neighbor classifier was applied andobtained an accuracy of 983 [14] The proposed methodconfirmed its superiority in the total accuracy compared tothe other systems The results which prove the superiority ofthe proposed method compared to the other existed systemsis demonstrated in Figure 8
ACCPrecision
SPECRecall
09091092093094095096097098099
1
1 2 3 4 5PE
RCEN
TAG
ENUMBER OF FEATURES
Figure 6 Performance measures of the original and compactfeatures set
STD
lt= 2873 gt 2873
0
0
ApEn
lt= 016 gt 016
Katz 1
1
lt= 146 gt 146
Figure 7 The classification rules obtained from the proposedsystem to classify the seizure-free (0) and the epileptic seizure (1)signals
Once the system detects the preictal phase the clinicreceives a notification about that patientThemain advantageof the hybrid system from the clinical point of view could besummarized as follows
(i) Classification and detecting of the epileptic seizuresand seizure-free signals from the EEG signal automat-ically
10 International Journal of Biomedical Imaging
90
931195 957 965
975 9775 9775 98 9813 983 99
Baja
j et a
l 20
13
Faus
t et a
l 20
10
Acha
rya e
t al
2009
Mar
tis et
al 2
013
Yuan
et al
201
1
Nag
hsh-
Nilc
hi et
al 2
010
Pach
ori e
t al
2014
Patid
ar et
al 2
017
Acha
rya e
t al
2012
Baja
j et a
l 20
12
Kum
ar et
al 2
015
e p
ropo
sed
met
hod
PROPOSED SYSTESMS USED TO FOR THE COMPARSION
8486889092949698
100
ACCU
RACY
IN P
ERCE
NTA
GE
()
Figure 8 The total accuracy () of various methods employed to detect the disorder of the epilepsy
(ii) The detection of the preictal phase from studying thehealthy and ictal phase of each patient
(iii) The detection of the preictal phase that provided theability to send warning alert to the physicians toprepare the medical assessment for the patient
(iv) The proposed method being robust and reliable as itsperformance was benchmarked using 10-fold cross-validation
(v) The dynamic behavior obtained from the fireflyalgorithm which made the system adaptive to manyfeatures according to each case study
(vi) A few sets of parameters required to analyze the EEGsignal typical three variables (119876 119903 119869)
The limitations of this research were summarized as follows
(i) The limited number of the studied subjects (typically100 per class)
(ii) The diagnosis process that may be reduced because ofdepending on some additional software prerequisites
5 Conclusion
In this paper an automated intelligent CAD tool has beenproposed to classify and detect epileptic seizures and seizure-free EEG signals This method provided an EEG signalanalysis using a hybrid data fusion method The data fusionmethod combined the collected features from two differentperspectives In the first perspective the EEG signal wasconsidered as an image Then the image was converted toa gray image and the GLCM was obtained to extract thetextures of the image such as contrast correlation power andhomogeneity In the second perspective the EEG signal wasdivided into smaller segments using the TQWT to extractthe time and frequency features A set of statistical nonlinear(chaotic) and power spectrum features were obtained from
each segment After the dataset was constructed a featurereduction algorithmbased on firefly optimizationwas used toreduce the irrelevant features and remove redundancy Thenan RFwas trained to classify and predict the epileptic seizuresand seizure-free EEG signals from the dataset The experi-mental results showed that the proposed method achieveda satisfactory degree of 99 accuracy 97 precision 97specificity 98 recall 98 F-measure and 95 MCC at thethird level of decomposition
Data Availability
The dataset used to support the findings of this research wasincluded in the article and can be found as a citation to ldquoRGAndrzejak K Lehnertz F Mormann C Rieke P DavidCE Elger Indications ofNonlinearDeterministic and Finite-Dimensional Structures in Time Series of Brain ElectricalActivity Dependence on Recording Region and Brain StatePhysical Review E 64 (2001) 061907rdquo
Conflicts of Interest
The authors declare that they have no conflicts of interest
References
[1] Epilepsy Fact Sheet 2018 httpwwwwhointmediacentrefactsheetsfs999en
[2] U R Acharya S V Sree G Swapna R J Martis and J S SurildquoAutomated EEG analysis of epilepsy a reviewrdquo Knowledge-Based Systems vol 45 pp 147ndash165 2013
[3] A Puce and M S Hamalainen ldquoA review of issues relatedto data acquisition and analysis in EEGMEG studiesrdquo BrainSciences vol 7 no 6 2017
[4] V Joshi R B Pachori and A Vijesh ldquoClassification of ictaland seizure-free EEG signals using fractional linear predictionrdquoBiomedical Signal Processing and Control vol 9 pp 1ndash5 2014
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
10 International Journal of Biomedical Imaging
90
931195 957 965
975 9775 9775 98 9813 983 99
Baja
j et a
l 20
13
Faus
t et a
l 20
10
Acha
rya e
t al
2009
Mar
tis et
al 2
013
Yuan
et al
201
1
Nag
hsh-
Nilc
hi et
al 2
010
Pach
ori e
t al
2014
Patid
ar et
al 2
017
Acha
rya e
t al
2012
Baja
j et a
l 20
12
Kum
ar et
al 2
015
e p
ropo
sed
met
hod
PROPOSED SYSTESMS USED TO FOR THE COMPARSION
8486889092949698
100
ACCU
RACY
IN P
ERCE
NTA
GE
()
Figure 8 The total accuracy () of various methods employed to detect the disorder of the epilepsy
(ii) The detection of the preictal phase from studying thehealthy and ictal phase of each patient
(iii) The detection of the preictal phase that provided theability to send warning alert to the physicians toprepare the medical assessment for the patient
(iv) The proposed method being robust and reliable as itsperformance was benchmarked using 10-fold cross-validation
(v) The dynamic behavior obtained from the fireflyalgorithm which made the system adaptive to manyfeatures according to each case study
(vi) A few sets of parameters required to analyze the EEGsignal typical three variables (119876 119903 119869)
The limitations of this research were summarized as follows
(i) The limited number of the studied subjects (typically100 per class)
(ii) The diagnosis process that may be reduced because ofdepending on some additional software prerequisites
5 Conclusion
In this paper an automated intelligent CAD tool has beenproposed to classify and detect epileptic seizures and seizure-free EEG signals This method provided an EEG signalanalysis using a hybrid data fusion method The data fusionmethod combined the collected features from two differentperspectives In the first perspective the EEG signal wasconsidered as an image Then the image was converted toa gray image and the GLCM was obtained to extract thetextures of the image such as contrast correlation power andhomogeneity In the second perspective the EEG signal wasdivided into smaller segments using the TQWT to extractthe time and frequency features A set of statistical nonlinear(chaotic) and power spectrum features were obtained from
each segment After the dataset was constructed a featurereduction algorithmbased on firefly optimizationwas used toreduce the irrelevant features and remove redundancy Thenan RFwas trained to classify and predict the epileptic seizuresand seizure-free EEG signals from the dataset The experi-mental results showed that the proposed method achieveda satisfactory degree of 99 accuracy 97 precision 97specificity 98 recall 98 F-measure and 95 MCC at thethird level of decomposition
Data Availability
The dataset used to support the findings of this research wasincluded in the article and can be found as a citation to ldquoRGAndrzejak K Lehnertz F Mormann C Rieke P DavidCE Elger Indications ofNonlinearDeterministic and Finite-Dimensional Structures in Time Series of Brain ElectricalActivity Dependence on Recording Region and Brain StatePhysical Review E 64 (2001) 061907rdquo
Conflicts of Interest
The authors declare that they have no conflicts of interest
References
[1] Epilepsy Fact Sheet 2018 httpwwwwhointmediacentrefactsheetsfs999en
[2] U R Acharya S V Sree G Swapna R J Martis and J S SurildquoAutomated EEG analysis of epilepsy a reviewrdquo Knowledge-Based Systems vol 45 pp 147ndash165 2013
[3] A Puce and M S Hamalainen ldquoA review of issues relatedto data acquisition and analysis in EEGMEG studiesrdquo BrainSciences vol 7 no 6 2017
[4] V Joshi R B Pachori and A Vijesh ldquoClassification of ictaland seizure-free EEG signals using fractional linear predictionrdquoBiomedical Signal Processing and Control vol 9 pp 1ndash5 2014
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
International Journal of Biomedical Imaging 11
[5] P Ghaderyan A Abbasi and M H Sedaaghi ldquoAn efficientseizure prediction method using KNN-based undersamplingand linear frequency measuresrdquo Journal of Neuroscience Meth-ods vol 232 pp 134ndash142 2014
[6] R B Pachori and S Patidar ldquoEpileptic seizure classificationin EEG signals using second-order difference plot of intrin-sic mode functionsrdquo Computer Methods and Programs inBiomedicine vol 113 no 2 pp 494ndash502 2014
[7] A R Hassan and A Subasi ldquoAutomatic identification ofepileptic seizures from EEG signals using linear programmingboostingrdquoComputerMethods and Programs in Biomedicine vol136 pp 65ndash77 2016
[8] M Sharma A Dhere R B Pachori and U R AcharyaldquoAn automatic detection of focal EEG signals using new classof timendashfrequency localized orthogonal wavelet filter banksrdquoKnowledge-Based Systems vol 118 pp 217ndash227 2017
[9] O Faust U R Acharya H Adeli and A Adeli ldquoWavelet-basedEEGprocessing for computer-aided seizure detection andepilepsy diagnosisrdquo Seizure vol 26 pp 56ndash64 2015
[10] Y Kumar M L Dewal and R S Anand ldquoRelative waveletenergy and wavelet entropy based epileptic brain signals clas-sificationrdquo Biomedical Engineering Letters vol 2 no 3 pp 147ndash157 2012
[11] R Sharma andR B Pachori ldquoClassification of epileptic seizuresin EEG signals based on phase space representation of intrinsicmode functionsrdquo Expert Systems with Applications vol 42 no3 pp 1106ndash1117 2015
[12] V Bajaj and R B Pachori ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1 pp17ndash21 2013
[13] Y Kaya M Uyar R Tekin and S Yıldırım ldquo1D-local binarypattern based feature extraction for classification of epilepticEEG signalsrdquo Applied Mathematics and Computation vol 243pp 209ndash219 2014
[14] T S Kumar V Kanhangad and R B Pachori ldquoClassificationof seizure and seizure-free EEG signals using local binarypatternsrdquo Biomedical Signal Processing and Control vol 15 pp33ndash40 2015
[15] M Mursalin Y Zhang Y Chen and N V Chawla ldquoAutomatedepileptic seizure detection using improved correlation-basedfeature selectionwith random forest classifierrdquoNeurocomputingvol 241 pp 204ndash214 2017
[16] L Wang W Xue Y Li et al ldquoAutomatic epileptic seizuredetection in EEG signals usingmulti-domain feature extractionand nonlinear analysisrdquo Entropy vol 19 no 6 2017
[17] Q Yuan W Zhou L Zhang et al ldquoEpileptic seizure detectionbased on imbalanced classification and wavelet packet trans-formrdquo Seizure vol 50 pp 99ndash108 2017
[18] S Lahmiri ldquoGeneralized Hurst exponent estimates differentiateEEG signals of healthy and epileptic patientsrdquo Physica AStatistical Mechanics and its Applications vol 490 pp 378ndash3852018
[19] U R Acharya F Molinari S V Sree S Chattopadhyay K HNg and J S Suri ldquoAutomated diagnosis of epileptic EEG usingentropiesrdquo Biomedical Signal Processing and Control vol 7 no4 pp 401ndash408 2012
[20] R Dhiman J S Saini and Priyanka ldquoGenetic algorithms tunedexpert model for detection of epileptic seizures from EEGsignaturesrdquo Applied Soft Computing vol 19 pp 8ndash17 2014
[21] A R Hassan S Siuly and Y Zhang ldquoEpileptic seizure detectionin EEG signals using tunable-Q factor wavelet transform andbootstrap aggregatingrdquo Computer Methods and Programs inBiomedicine vol 137 pp 247ndash259 2016
[22] S Patidar and T Panigrahi ldquoDetection of epileptic seizure usingKraskov entropy applied on tunable-Q wavelet transform ofEEG signalsrdquo Biomedical Signal Processing and Control vol 34pp 74ndash80 2017
[23] J Jia B Goparaju J Song R Zhang and M B WestoverldquoAutomated identification of epileptic seizures in EEG signalsbased on phase space representation and statistical features inthe CEEMDdomainrdquo Biomedical Signal Processing and Controlvol 38 pp 148ndash157 2017
[24] T Gandhi B K Panigrahi and S Anand ldquoA comparative studyof wavelet families for EEG signal classificationrdquoNeurocomput-ing vol 74 no 17 pp 3051ndash3057 2011
[25] IW Selesnick ldquoWavelet transformwith tunableQ-factorrdquo IEEETransactions on Signal Processing vol 59 no 8 pp 3560ndash35752011
[26] S Patidar and R B Pachori ldquoClassification of cardiac soundsignals using constrained tunable-Q wavelet transformrdquo ExpertSystems with Applications vol 41 no 16 pp 7161ndash7170 2014
[27] S Patidar R B Pachori AUpadhyay andU RajendraAcharyaldquoAn integrated alcoholic index using tunable-Q wavelet trans-form based features extracted from EEG signals for diagnosisof alcoholismrdquo Applied Soft Computing vol 50 pp 71ndash78 2017
[28] R G Andrzejak K Lehnertz F Mormann C Rieke P Davidand C E Elger ldquoIndications of nonlinear deterministic andfinite-dimensional structures in time series of brain electricalactivity dependence on recording region and brain staterdquoPhysical Review E Statistical Nonlinear and SoftMatter Physicsvol 64 no 6 2001
[29] P Swami A K Godiyal J Santhosh B K Panigrahi M Bhatiaand S Anand ldquoRobust expert system design for automateddetection of epileptic seizures using SVM classifierrdquo in Proceed-ings of the 2014 3rd IEEE International Conference on ParallelDistributed and Grid Computing PDGC 2014 pp 219ndash222India December 2014
[30] P Swami T K Gandhi B K Panigrahi M Tripathi and SAnand ldquoA novel robust diagnostic model to detect seizures inelectroencephalographyrdquo Expert Systems with Applications vol56 pp 116ndash130 2016
[31] H Ocak ldquoAutomatic detection of epileptic seizures in EEGusing discrete wavelet transform and approximate entropyrdquoExpert Systems with Applications vol 36 no 2 pp 2027ndash20362009
[32] J-H Kang Y G Chung and S-P Kim ldquoAn efficient detectionof epileptic seizure by differentiation and spectral analysis ofelectroencephalogramsrdquo Computers in Biology and Medicinevol 66 pp 352ndash356 2015
[33] S Barua M U Ahmed C Ahlstrom S Begum and P FunkldquoAutomated EEG Artifact Handling with Application in DriverMonitoringrdquo IEEE Journal of Biomedical andHealth Informatics2017
[34] G E Polychronaki P Y Ktonas S Gatzonis et al ldquoComparisonof fractal dimension estimation algorithms for epileptic seizureonset detectionrdquo Journal of Neural Engineering vol 7 no 42010
[35] J Gorecka ldquoDetection of ocular artifacts in EEG data usingthe Hurst exponentrdquo in Proceedings of the 20th InternationalConference onMethods andModels in Automation and RoboticsMMAR 2015 pp 931ndash933 August 2015
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
12 International Journal of Biomedical Imaging
[36] F Parastesh Karegar A Fallah and S Rashidi ldquoECG basedhuman authenticationwith usingGeneralizedHurst Exponentrdquoin Proceedings of the 25th Iranian Conference on ElectricalEngineering ICEE 2017 pp 34ndash38 May 2017
[37] I Fister I Fister Jr X-S Yang and J Brest ldquoA comprehensivereview of firefly algorithmsrdquo Swarm and Evolutionary Compu-tation vol 13 no 1 pp 34ndash46 2013
[38] L Zhang K Mistry C P Lim and S C Neoh ldquoFeature selec-tion using firefly optimization for classification and regressionmodelsrdquo Decision Support Systems vol 106 pp 64ndash85 2018
[39] L Breiman ldquoRandom forestsrdquoMachine Learning vol 45 no 1pp 5ndash32 2001
[40] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006
[41] J P Kandhasamy and S Balamurali ldquoPerformance analysisof classifier models to predict diabetes mellitusrdquo ProcediaComputer Science vol 47 pp 45ndash51 2015
[42] G Varoquaux ldquoCross-validation failure Small sample sizes leadto large error barsrdquo NeuroImage 2017
[43] O Faust U R Acharya L C Min and B H C Sputh ldquoAuto-matic identification of epileptic and background eeg signalsusing frequency domain parametersrdquo International Journal ofNeural Systems vol 20 no 2 pp 159ndash176 2010
[44] U R Acharya C K Chua T-C Lim Dorithy and J SSuri ldquoAutomatic identification of epileptic EEG signals usingnonlinear parametersrdquo Journal of Mechanics in Medicine andBiology vol 9 no 4 pp 539ndash553 2009
[45] R J Martis U R Acharya J H Tan et al ldquoApplication ofintrinsic time-scale decomposition (ITD) to EEG signals forautomated seizure predictionrdquo International Journal of NeuralSystems vol 23 no 5 pp 1557ndash1565 2013
[46] Q YuanWZhou S Li andDCai ldquoEpileptic EEG classificationbased on extreme learning machine and nonlinear featuresrdquoEpilepsy Research vol 96 no 1-2 pp 29ndash38 2011
[47] A R Naghsh-Nilchi and M Aghashahi ldquoEpilepsy seizuredetection using eigen-system spectral estimation and MultipleLayer Perceptron neural networkrdquo Biomedical Signal Processingand Control vol 5 no 2 pp 147ndash157 2010
[48] V Bajaj and R B Pachori ldquoClassification of seizure andnonseizure EEG signals using empirical mode decompositionrdquoIEEE Transactions on Information Technology in Biomedicinevol 16 no 6 pp 1135ndash1142 2012
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom
International Journal of
AerospaceEngineeringHindawiwwwhindawicom Volume 2018
RoboticsJournal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Active and Passive Electronic Components
VLSI Design
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Shock and Vibration
Hindawiwwwhindawicom Volume 2018
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawiwwwhindawicom
Volume 2018
Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom
The Scientific World Journal
Volume 2018
Control Scienceand Engineering
Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom
Journal ofEngineeringVolume 2018
SensorsJournal of
Hindawiwwwhindawicom Volume 2018
International Journal of
RotatingMachinery
Hindawiwwwhindawicom Volume 2018
Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawiwwwhindawicom Volume 2018
Hindawiwwwhindawicom Volume 2018
Navigation and Observation
International Journal of
Hindawi
wwwhindawicom Volume 2018
Advances in
Multimedia
Submit your manuscripts atwwwhindawicom