a pervasive approach to eeg-based depression detection€¦ · complexity psychiatrist....

Research ArticleA Pervasive Approach to EEG-Based Depression Detection

Hanshu Cai1 Jiashuo Han1 Yunfei Chen1 Xiaocong Sha1 Ziyang Wang1 Bin Hu 123

Jing Yang4 Lei Feng5 Zhijie Ding6 Yiqiang Chen7 and Juumlrg Gutknecht8

1Gansu Provincial Key Laboratory of Wearable Computing School of Information Science and EngineeringLanzhou University Lanzhou China2CAS Center for Excellence in Brain Science and Intelligence Technology Shanghai Institutes for Biological SciencesChinese Academy of Sciences Shanghai China3Beijing Institute for Brain Disorders Capital Medical University Beijing China4Department of Child Psychology Lanzhou University Second Hospital Lanzhou China5Beijing Anding Hospital Capital Medical University Beijing China6TheThird Peoplersquos Hospital of Tianshui City Tianshui China7Institute of Computing Technology Chinese Academy of Sciences Beijing China8Computer Systems Institute ETH Zurich Zurich Switzerland

Correspondence should be addressed to Bin Hu bhlzueducn

Received 31 March 2017 Revised 17 November 2017 Accepted 4 January 2018 Published 6 February 2018

Academic Editor Haiying Wang

Copyright copy 2018 Hanshu Cai et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Nowadays depression is the worldrsquos major health concern and economic burden worldwide However due to the limitations ofcurrentmethods for depression diagnosis a pervasive andobjective approach is essential In the present study a psychophysiologicaldatabase containing 213 (92 depressed patients and 121 normal controls) subjects was constructed The electroencephalogram(EEG) signals of all participants under resting state and sound stimulation were collected using a pervasive prefrontal-lobe three-electrode EEG system at Fp1 Fp2 and Fpz electrode sites After denoising using the Finite Impulse Response filter combining theKalman derivation formula Discrete Wavelet Transformation and an Adaptive Predictor Filter a total of 270 linear and nonlinearfeatures were extractedThen theminimal-redundancy-maximal-relevance feature selection technique reduced the dimensionalityof the feature space Four classification methods (Support Vector Machine K-Nearest Neighbor Classification Trees and ArtificialNeural Network) distinguished the depressed participants fromnormal controlsThe classifiersrsquo performances were evaluated using10-fold cross-validation The results showed that K-Nearest Neighbor (KNN) had the highest accuracy of 7927 The result alsosuggested that the absolute power of the theta wave might be a valid characteristic for discriminating depressionThis study provesthe feasibility of a pervasive three-electrode EEG acquisition system for depression diagnosis

1 Introduction

Depression is a common mood disorder which might causepersistent feeling of sadness loss of interest and impairmentof memory and concentration Depressed patients normallyexperience cognitive impairment and suffer long and severeemotional depression In severe cases some patients willexperience paranoia and illusion [1] According to the WorldHealth Organization statistics gt300 million individualssuffer from depression worldwide approximately 800000

people die due to it every year [2] Thus depression ispredicted to become the second most common disease afterheart disease by the year 2020 [3] Hence the diagnosis ofdepression in the early curable stages is critical andmight savethe life of a patient [4]

Presently the study on the human cerebral is currentlyunder intensive focus in order to understand the mecha-nism underlying persistent negative emotion and depressionTherefore the most commonly used diagnosis of depressionis a scale-based interview conducted by a psychologist or

HindawiComplexityVolume 2018 Article ID 5238028 13 pageshttpsdoiorg10115520185238028

2 Complexity

psychiatrist The current international standard mostly usedis ldquoIn Diagnostic and Statistical Manual of Mental Disorders(Fourth Edition)rdquo (DSM-IV) [5] and the clinical test Mini-Mental State Examination (MMSE) is commonly applied[6] Other conventional psychometric questionnaires such asBeck depression inventory (BDI) [7] and Hamilton Depres-sion Rating Scale (HDRS) [8] are also used as screening toolsrather than as the instrument for the diagnosis of depression

The current methods of depression detection are human-intensive and the results are dependent on the doctorrsquosexperience Furthermore depressed individuals are less likelyto seek help due to fear of stigma and the nature of thedisorder As a result a large number of depressed patients notdiagnosed accurately do not receive optimal treatment andadequate recovery period Therefore finding convenient andeffectivemethods for the detection of depression is an emerg-ing topic for research With the latest advances in the sensorand mobile technology the exploration using physiologicaldata for the diagnosis of mental disorder opens a new avenuefor an objective and accurate tool for depression detectionAmong all kinds of physiological data electroencephalogram(EEG) reflects emotional human brain activity in real time[9]

The EEG signal is a recording of the spontaneousrhythmic electrical activity of brain neurons from the scalpsurface Since the earliest discovery from the rabbit andmonkey brain and the first recording of the human EEGsignal by German psychiatrist Hans Berger in 1926 studieson the analytical method of EEG and the interpretation of theassociation between the brain function and mental disordershave been continued for over a century [10] Neurosciencepsychology and cognitive science research showed that amajority of the psychological activities and cognitive behav-ior could be indicated by EEG [11ndash13] The EEG signal isclosely related to the brain activities and emotional statesand it could reflect the emotional transformation in real timeCole and Ray [14] found that the EEG signal collected fromthe parietal lobe of brain is associated with the cognitive tasksand emotional states Klimesch et al found that the alphawaves with low frequency could reflect some of the features ofattention such as vigilance and expectations [15] Srinivasanet al demonstrated that the frequency domain features ofEEG could be used to predict the level of attention [16]Therefore the EEG signal is critical for understanding theprocessing of human brain information and emotional statetransformation

The studies on EEG could be used to understand themechanism underlying brain activity human cognitive pro-cess and diagnosis of brain disease as well as the field ofthe Brain Computer Interface (BCI) which has attractedmuch attention in recent years [17] Compared to Com-puted Tomography (CT) and functionalMagnetic ResonanceImaging (fMRI) EEG has a higher time resolution a lowermaintenance cost and a simpler operation method Thusas an objective physiological method to obtain data EEGwas proposed as a nonintrusive approach to study cogni-tive behavior [18ndash20] and other illness symptoms such asinsomnia [21ndash23] epilepsy [24ndash26] and sleep disorder [27]EEG has also been used in the diagnosis of mental disorders

such as anxiety [28ndash30] psychosis [31ndash34] and depression[35ndash38] In addition depression as a mental disorder withclinical manifestations such as significant depression andslow thinking is always accompanied by abnormal brainactivity and obvious emotional alternation Therefore as amethod tracking the brain functions EEG can detect theseabnormal activities

The frequency of the EEG signal can be divided into 5wave-bands delta wave (lt4Hz) which normally appears inan adultrsquos slow-wave sleep theta wave (4ndash8Hz) which isusually found when someone is sleepy alpha wave (8ndash14Hz)which is normally detected when someone is relaxed betawave (14ndash30Hz) which commonly appears when someone isactively thinking and gamma wave (30ndash50Hz) which couldappear during meditationThe EEG signals undergo changesin the amplitude as well as frequency while different mentaltasks are performed [39ndash42]

Presently for research purposes the most commonlyused are 128-electrode and 256-electrode EEG systems [4344] which are specifically designed for research purposesThe operation of the instruments was not only difficult toinitiate but also it required technicians to apply conductivegel to each electrode on the participantrsquos head before each useThe preparation process alone takes 30 minutes on averageIn addition these EEG systems are expensive Overall thesesystems are not practical for pervasive depression detec-tion

In the present study the pervasive three-electrode EEGacquisition system developed independently by the Ubiq-uitous Awareness and Intelligent Solutions Lab (UAIS) ofLanzhou University [45] was employed to construct adatabase containing both depressed patients and normalcontrols Thus the use of the latest data processing tech-nique and machine learning to explore a pervasive EEG-based depression detection system has been the focus ofinvestigation In order to support this research

(1) A pervasive three-electrode EEG acquisition systemhas been introduced (Section 21)

(2) A psychophysiological experiment has been con-ducted in which EEG of 213 participants has beenrecorded These physiological data provided a com-prehensive database for further analysis construc-tion and evaluation of a pervasive EEG-based depres-sion detection system (Sections 22 and 23)

(3) Several EEG preprocessing steps and methods wereapplied on the raw EEG data (Section 31)

(4) 270 features were identified and extracted from therecoded database By employing a feature selectiontechnique an optimum feature matrix was con-structed for the depression classification process (Sec-tion 32)

(5) Four classification algorithms including K-NearestNeighbor (KNN) Support Vector Machine (SVM)Classification Tree (CT) and Artificial Neural Net-work (ANN) have been evaluated and comparedusing a 10-fold cross-validation (Section 4)

Complexity 3

C4CzC3

F3F7

Fp1 Fp2

Fz F4 F8

T3

T5 T6

O2O1

T4

P4PzP3

A2A1

Nasion

Inion

Figure 1 The international 10-20 system

2 Pervasive Three-Electrode EEGDatabase Construction

21 Pervasive Three-Electrode EEG Acquisition System The10-20 system proposed by Jasper in 1958 defined the nameof the electrode and later became the international stan-dard EEG placement system [46] With the development ofsensor technology the electrode became smaller than thatin previous systems and the electrodes recorded a detailedEEG In 1985 Chatrian et al added extra electrodes inintermediate sites halfway between those of the existing 10-20system thereby expanding it to a 64-electrode system [47]Due to the complexity of the full-brain 128-electrode and256-electrode systems the investigators restricted themselvesfrommobile and pervasive application Thus with the devel-opment of universal and pervasive electronic technology the8-electrode and 16-electrode systems with small volume werealso developed gradually

As shown in Figure 1 F represents the frontal lobeT represents the temporal lobe C represents the center Prepresents the parietal lobe and O represents the occipitallobe EEG reacts to the biological activity of the brain tissuethereby indicating the functional status of the brain [48]TheEEG signal collected from the different locations of the scalpreflects a variety of information For example EEG from thefrontal lobe reflects human memory computational powerattention and responsiveness EEG from the parietal lobe isassociated with somatic responses EEG from the occipitallobe can be used as a reference for visual reactions EEGfrom temporal lobe is related to auditory reactionsThereforefor different research direction and purpose the appropriateEEG collection location is essential

Prefrontal cortex is the center of consciousness thus thebetter the control of the forehead cortex the better the emo-tional control Jasper studied the resting-state EEG of severedepression patients showing that when the body sufferedfrom severe depression the activity of the cerebral cortexwas altered [49] Nauta emphasized that the prefrontal cortexplayed amajor role in different aspects of emotional processes

[50] Rolls put forward the importance of prefrontal cortexfor emotional andmotivational processes [51]Harmon-Jonessuggested that the specific forms of anger or anger elicited inparticular contexts are associated with left-sided prefrontalactivation [52] In conclusion the above studies have shownthat the electrode sites located in the prefrontal cortex areassociated with emotional process and psychiatric disordersTherefore Fp1 Fp2 and Fpz are the ideal choices of scalpposition in the current experiment The hair in the frontallobe is absent and contact dry electrode should be sufficientwithout the need for applying conductive gel The pervasivethree-electrode EEG acquisition system (Figure 2) developedby UAIS from Lanzhou University [53] runs on rechargeablebattery and transmits all the EEG data through Bluetooth20 wirelessly The system is extremely small in size and canbe easily placed on the location The sampling frequency is250Hz and according to the EGI engineers all electrodeshave an impedance of lt50 kΩ Since the frequency of EEG is05ndash50Hz the passband of the EEG acquisition is 05ndash50Hz

22 Experiment Method Compared to the normal controlsdepressed patients responded differently to outside stimulus[54 55] The feedback of the depressed patients to thepositive and negative stimuli weakened As the positivestimulus feedback weakened further the overall performancewas negative emotions and reflected as such in the emo-tional response of the different subsystems In summaryno significant difference was observed in the positive stim-ulus between normal controls and depressed patients anddepressed patients would produce more negative emotionsunder negative stimulus as compared to normal controlsBeckrsquos cognitive behavioral model of depression postulatedthat the depressed patients are likely to support a negativeview of themselves the world and even the future In orderto maintain this negative self-view they even resist theenvironmental feedback that is inconsistent with the view[56] Epstein et al suggested that in comparison to normalcontrols depressed patients responded with less bilateralventral striatal activation to positive stimuli which leadsto the decreased interest in performance of activities [57]Bylsma et al proved that depressed patients exhibit lessreactivity to all stimuli and events irrespective of positive ornegative nature [58]

Therefore recording and analysis of the EEG signal indifferent stimuli may help in the identification of patientswith depression This study was designed to record theparticipantsrsquo EEG in four different cases in resting stateunder negative stimulus under neutral stimulus and underpositive stimulusThe source of stimulus is soundtracks fromthe International Affective Digitized Sounds (IADS-2) [59]which is a standardized database of 167 naturally occurringsounds widely used in the study of emotions

The experiment was performed in a quiet room Firstlythe experiment objective and procedures were described tothe participants Then the pervasive three-electrode EEGacquisition system was placed on the participantsrsquo foreheadsand checked for reception After one minute of relaxationthe experiment begins again At first stage 90 s of resting-state EEG was recorded Then the participants were asked to

4 Complexity

EEG signal

Ground

FP2 FPZ FP1

Figure 2 Pervasive three-electrode EEG acquisition system using Fp1 Fp2 and Fpz positions

Described experimentobjective and proceduresto the participants

Place the acquisitionsystem on participantsforehead and check forreception

Experiment completes

Play the sixthstimulation soundtrackto the participants andhave a 6 s break

Play the first stimulationsoundtrack to theparticipants and have a6 s break

break state EEG and have a 6 sRecord 90 s of resting-

Figure 3 Process of EEG acquisition

Table 1 Audio stimulation profile

Number Name Property(1) Cattle Neutral(2) Painting Neutral(3) Babies cry Negative(4) Dentist drill Negative(5) Baby Positive(6) Crowd Positive

remain seated with eyes closed with as little body movementsas possible followed by another minute of rest In the secondstage stimulation soundtracks will be played to participantsEach soundtrack was 6 s long with a 6 s break between eachsoundtrackThe process would continue until the experimentis completed The process of EEG acquisition is shown inFigure 3

A total of 6 stimulation soundtracks (according to IADS-2) existed including 2 neutral stimulation soundtracks 2negative stimulation soundtracks and 2 positive stimulationsoundtracks Table 1 describes each audio stimulation

23 Psychophysiological Database Of the total 250 partici-pants 213 (92 depressed patients and 121 normal controls)completed the experiment successfully The raw EEG data

from all the electrodes were recorded Depressed participantswere selected by professional psychiatrists using MMSE [6]which is a 30-point questionnaire used by the psychiatristduring a face-to-face interview to assess the degree of cog-nitive dysfunction in patients with diffuse brain disorders Inaddition all participants are asked to fill the following scalesfor cross-referencing

(A) The Patient Health Questionnaire (PHQ-9) [60]is a 9-question-based multipurpose instrument forscreening diagnosing monitoring and measuringthe severity of depression We chose this question-naire in order to find the relevance between the EEGcharacteristic and the severity of depression

(B) Life Event Scale (LES) [61] contains 48 questionsincluding events of family work and social supportThe influence of each event is evaluated for severityduration and frequency We chose this questionnairefor cross-referencing purposes

(C) Pittsburgh Sleep Quality Index (PSQI) [62] contains19 self-reported items creating 7 components todiagnose sleep disorders We chose this index toexplore the direct link between sleep qualities withdepression in EEG

(D) Generalized Anxiety Disorder Scale-7 (GAD-7) [63]contains only 7 self-report questions for screening

Complexity 5

and measuring the severity of generalized anxietydisorder We chose this questionnaire for cross-referencing between depression and anxiety

3 Data Processing

In this study all preprocessing and data analyses have beenimplemented using MATLAB software (version R2014a)

31 Preprocessing EEG is a noninvasive method of captur-ing the physiological signal of brainwave activity HoweverEEG data recorded are normally mixed with interferencesfrom surrounding environment such as close-by power lineFurthermore other physiological signals including elec-trocardiogram (ECG) electrooculogram (EOG) and elec-tromyograph (EMG) could also be detected and recordedby EEG sensors [55] To ensure an accurate result in thefeature selection and classification all the raw data should bedenoised first

ECG is a smooth signal among the physiological electricalsignals with a large amplitude As the heart is located distallyfrom the head the ECG signal will be greatly attenuatedwhenspread to the scalp EMG is produced by muscle contractionwith an amplitude of 10120583V to 15mV The frequency of EMGis concentrated primarily in the high band gt 100Hz Power-line interference focuses on fixed operation frequency Inorder to remove these interference signals we followed theresults of several investigators Yang proposed a cascade ofthree adaptive filters based on the least mean squares (LMS)algorithm and verified that the proposed filter reduced theinterference in EEG signals [64] Tong et al validated theuse of independent component analysis (ICA) for an efficientsuppression of the interference of ECG from EEG [65] TheNational Institute of Mental Health announced that using anadaptive filter to estimate the contaminants can subtract themfrom the EEG data [66]

No overlap occurred between the frequency of EEG signaland power-line interferences EMG and ECG thus FiniteImpulse Response (FIR) filter based on the Blackman timewindow was used to remove these interference signals Theadequate linearity of the FIR filter is widely used in modernelectronic communication It can guarantee any amplitudefrequency characteristics simultaneously with strict linearphase-frequency characteristics In addition the unit sam-pling response is finite which stabilized the filter In order toreduce the energy leakage of the spectrum the signal can betruncated by different interception functionsThis truncationfunction is known as the window functionThe time domainrepresentation of the Blackman time window is

119908 (119899) = [042 minus 05 cos( 2119898119873 minus 1) + 008 cos(

4119898119873 minus 1)]

sdot 119877119873 (119899) (1)

where119877119873(119899) is the rectangular window function and119873 is thelength of truncated data

The resulting EEG signal is retained only between fre-quencies in the range of 05ndash50Hz However the frequencyof EOG overlaps within this range Although all participants

were asked to remain seated with eyes closed their EOGwas recorded inevitably while using the prefrontal-lobe EEGsites such as Fp1 Fp2 and Fpz A general model for EOGcontamination can be described by

119910 (119899) = 119909 (119899) + 119865 (119903) (2)

where 119910(119899) and 119909(119899) are the samples of the recorded (includ-ing noise) and true EEG respectively 119903 represents the sourceEOG and 119865 is an unknown transfer function

Kalman filter is an optimal recursive data processingalgorithm which has been widely utilized in several appli-cations such as industrial control systems radar targettracking communications and signal processing aeroenginediagnosis and intelligent robots Kalman filter is based onthe previous estimated value and the observed value of thecurrent time to estimate the current value of the statedvariable Thus the frequency of the EOG artifact wouldnot exceed 15Hz and the approximate EOG signal and theamplitude of the brain in the low frequency band are smallAs a result the Kalman derivation formula combines theDiscrete Wavelet Transformation (DWT) and an AdaptivePredictor Filter (APF) to estimate the pure EOG artifact

The denoising model proposed in the present studyinvolves the following steps (1) signal decomposition (2)ocular artifacts (OA) zones detection (3) signal predictionand (4) signal reconstruction Herein DWT was used todecompose the EEG signals and detect the OZ zones Thefrequency range of the EEG signal was 0ndash64Hz while theOAoccurred in 0ndash16Hz The multiscale DWT decompositionwas used to extract the low frequency components and non-stationary time series which were then divided into severalapproximate stationary time series Thus the conventionalforecasting methods such as Kalman filter can predict theshape of the true wave of decomposition signals accuratelySubsequently the Adaptive Auto Regressive (AAR) modelsand an Adaptive Predictor Filter (APF) were applied toimprove the prediction The APF uses an adaptive filter toestimate the future values of signals based on their past valuesFinally the EOG artifacts were removed from the raw EEGsignal and the data were ready for further processing

32 Features Matrix Construction The features matrix con-sists of 119899 rows and 119898 columns where 119899 represents thenumber of EEG data and119898 represents the number of featuresextracted from each EEG The present study constructed thetraining effective features matrix using three steps as follows

(1) Identify and extract all the efficient features for each setof EEG data such that each row represents a feature vector

(2) Each row of the features matrix is selected by featureselection that is the most suitable feature is selected from allthe extracted features to form a final feature vector

(3) Each row of the feature vectors is tagged by depressionor nondepression

321 Feature Extraction The EEG signal presents weaknonlinear and time-sensitive characteristic which exhibitstypically complex dynamics The feature of EEG will changewith the emotional state transformation The analysis of

6 Complexity

EEG data displayed different linear features such as peakvariance and skewness that were used in recent literature[67ndash70] Efforts have been made in determining nonlinearparameters such as Correlation Dimension for pathologicalsignals which are shown as useful indicators of pathologies[71] In order to obtain the feature matrix we must firstperform the feature extraction of the pretreated EEG TheEEG features are mainly divided into Time Domain Featuresand Frequency Domain Features Owing to the nonlinearityand randomness of the EEG signal this study extracts thenonlinear features such as the Correlation Dimension andShannon Entropy in addition to the above EEG featuresFinally the following features were selected for extraction

(1) Time Domain Features Time domain constitutes themost intuitive EEG features The EEG signals are collectedat a certain time and frequency The artifacts are directlyremoved from the time domain EEG signal and usefulinformation was extracted as a time domain feature thatcan be used for continuous prolonged EEG detection Thetime domain features extracted in this study include peakvariance skewness kurtosis and Hjorth parameter Hjorthparameters are indicators of statistical properties used insignal processing in the time domain introduced by Hjorthin 1970 [72] the parameters include activity mobility andcomplexity Among them the activity parameters representthe signal power and the variance of time function Themobility parameters represent the mean frequency or theproportion of standard deviation of the power spectrumThecomplexity parameters represent the change in frequencyThese parameters are usually used to analyze the EEG signalsfor feature extraction

(2) Frequency Domain Features Frequency domain is a toolfor characterizing and classifying the EEG signals Herein thefrequency domain features are relative centroid frequencyabsolute centroid frequency relative power and absolutepower

(3) Nonlinear Features The EEG signals are nonstationaryand random they also include some of the characteristicsof the nonlinear dynamics system With increasing numberof studies on the EEG signals the nonlinearity has beenunder intensive focus worldwide Therefore processing andanalyzing the EEG signal based on the nonlinear dynamicstheory become a new research direction The nonlinearfeatures extracted in this study include 1198620-complexity Kol-mogorov Entropy Shannon Entropy CorrelationDimensionand Power-Spectral Entropy

(A) The 1198620-complexity was proposed by Shen et al [73]to resolve the issue of over-coarse graining prepro-cessing in Lempel-Ziv complexity (LZC) [74] Thecore of the algorithm is to decompose the sequenceinto regular and irregular components and the 1198620-complexity defines the proportion of irregularities inthe sequence The greater the proportion of its sharethe closer the time domain signal to the randomsequence and thus the greater the complexity The

doctrine presumes that a signal can be divided intoregular part and stochastic components If 1198600 is ameasurement of the signal and1198601 is themeasurementcorresponding to the stochastic part 1198620-complexityis defined as the ratio of 1198601 and 1198600 Supposedly theEEG signal to be analyzed is 119909(119899) 119899 = 0 1 119873minus1with a length of 119873 samples then the 1198620-complexitycan be calculated with the power spectra as followsThe fast Fourier transform (FFT) of the signal is asfollows

119883 (119896) = 1119873119873minus1sum119899=0

119909 (119899) 119890minus119895(2119896120587119899119873) 119896 = 0 1 119873 minus 1 (3)

Themean amplitude of the power spectrum119883(119896) is asfollows

119872 = 1119873119873minus1sum119896=0

|119883 (119896)|2 (4)

119883(119896) less than 119872 are replaced by 0 to obtain a newspectrum series 119884(119896)

119884 (119896) = 119883(119896) |119883 (119896)|2 gt 1198720 |119883 (119896)|2 le 119872 (5)

The inverse FFT (IFFT) of 119884(119896) is as follows119910 (119899) = 119873minus1sum

119896=0

119884 (119896) 119890119895(2119896120587119899119873) 119899 = 0 1 119873 minus 1 (6)

The power of stochastic part 1198601 is extracted and the1198620-complexity was estimated

1198601 =119873minus1sum119899=0

1003816100381610038161003816119909 (119899) minus 119910 (119899)10038161003816100381610038162

1198600 =119873minus1sum119899=0

|119909 (119899)|2

1198620 = 11986011198600

(7)

(B) Kolmogorov Entropy was used to measure the rateof loss of information per unit of time Positive andfinite entropy represents that the time series and thedynamic underlying phenomenon are chaotic Zeroentropy indicates a regular phenomenon in the spacephase Infinite entropy refers to a stochastic and non-deterministic phenomenon Kolmogorov Entropy isdefined as the average rate of loss of information asfollows

KE = minuslim120591rarr0

lim120576rarr0

lim119899rarrinfin

1119899120591 sum1198940 sdotsdotsdot119894119899

1198751198940 sdotsdotsdot119894119899minus1 ln1198751198940 sdotsdotsdot119894119899minus1 (8)

Complexity 7

(C) Shannon Entropy was introduced by Shannon in 1948in an article entitled ldquoA Mathematical Theory ofCommunicationrdquo [75] The size of the informationof a message is directly related to its uncertaintyThe amount of information is equal to the amountof uncertainty Shannon Entropy is a measure ofuncertainty of a randomvariable and a randomsignalThe larger the entropy the greater the uncertainty andrandomness In the present study the entropy usedto process EEG can be viewed as a measure of theorder in the signal which measures the skewness anduncertainty [76] In the case of random variables withknown probability distribution the entropy is definedby

119867(119883) = sum119909isin120594

119901 (119909) log119901 (119909) (9)

where 119883 is a random variable with probability distri-bution 119901(119909) and alphabet set 120594 [77]

(D) Correlation Dimension indicates the dynamic fea-tures of the EEG signal The greater the Correla-tion Dimension number the complicated the EEGtime series The Correlation Dimension is a fractaldimension often computed from the time seriesillustration It is a simplified phase space diagramconstructed from a single data vectorThe fundamen-tal Correlation Dimension algorithm was introducedby Grassberger and Procacia in 1983 [5] and can beexpressed as below

CD = lim119903rarr0

( ln119862 (119903)ln 119903 ) (10)

where 119862(119903) is the correlation integral and 119903 is theradial distance around each reference point

(E) Power-Spectral Entropy is a sequence of powerdensity with the frequency distribution obtainedby Fourier transform The calculated entropy ofthe power spectrum (referred to as Power-SpectralEntropy) can be implemented easily The Power-Spectral Entropy is used to analyze the timing signalsin EEG data The entropy can be used as a physicalindicator to estimate the quality and intensity of brainactivity The larger the entropy the more active thebrain

All linear and nonlinear features (Table 2) were extractedfrom alpha wave beta wave delta wave theta wave gammawave and full-band EEG of each electrode (Fp1 Fp2 andFpz) Therefore a total of 270 features (15 basic features times 6frequencies times 3 electrodes) were extracted All the involvedlinear and nonlinear features are common information aboutEEG

322 Feature Selection Feature Selection is used to select arelevant subset of all available features which not only yieldsa small dimensionality of the classification problem but also

Table 2 Features used in the feature extraction process

Name PropertyCentroid frequency

Linear features

Relative centroid frequencyAbsolute centroid frequencyRelative powerAbsolute powerPeakVarianceSkewnessKurtosisHjorthPower-Spectrum Entropy

Nonlinear featuresShannon EntropyCorrelation DimensionC0-complexityKolmogorov Entropy

reduces the noise (irrelevant features) We further deducedthe types of features suitable for suppressing the EEG signalrecognition by inspecting the features selected by the appliedalgorithm

The feature evaluation function focuses on the relationbetween the features and the target class which tends toinvolve redundant features influencing the learning accuracyand results In order to achieve these results we appliedtheminimal-redundancy-maximal-relevance (MRMR) tech-nique to perform the feature selection The MRMR featureselection criterion was proposed by Peng et al [78] in orderto resolve the issue by evaluating both feature redundancyand relevance simultaneously in particular max-relevancedenoted as max119863(119878 119888) refers to maximizing the relevance ofa feature subset 119878 to the class label 119888 In [1] the relevance of afeature subset is defined as

max119863 (119878 119888) = 1|119878| sum119891119894isin119878

Φ(119891119894 119888) (11)

where Φ(119891119894 119888) denotes the relevance of a feature 119891119894 to 119888 Φcould be estimated using any correlation measures

Feature redundancy is defined based on the pairwisefeature dependence If two relevant features highly dependon each other the class-discrimination power would notchange dramatically if one of the features was removed Min-redundancy min119877(119878) is used to select a feature subset ofmutually exclusively features The redundancy of a featuresubset is defined as

min119877 (119878) = 1|119878|2 sum119891119894 119891119895isin119878

Φ(119891119894 119891119895) (12)

MRMR is defined as the simple operator maximizing 119863 andminimizing 119877 consecutively In [1] the incremental searchmethod was used to find the near-optimal features The

8 Complexity

feature subset 119878119898minus1 of119898minus1 selected feature is utilized to selectthe119898-order feature that optimizes the following criterion

max119891119895notin119878119898minus1

[[Φ (119891119894 119888) minus 1

119898 minus 1 sum119891119894isin119878119898minus1

Φ(119891119894 119891119895)]] (13)

323 Effective Tagging Each feature vector (each row of thefeature matrix) has to be marked with a specific emotionaltag In this study we divided the experimental populationinto two categories depressed patients and normal controlsAll eigenvectors are tagged as depressed and nondepressed

4 Classification

SVM KNN and CT are the widely used classificationalgorithms in the majority of the EEG-related studies Inthe present study we evaluated the performance of theseclassifiers (SVM KNN and CT) plus the Artificial NeuralNetwork (ANN) classifier in the depression detection pro-cess All classifications and 10-fold cross-validations havebeen implemented using the MATLAB software (versionR2014a)

41 Classification Techniques

411 SVM SVM proposed by Cortes and Vapnik [79] in1995 is a supervised learning model and regression methodIt exhibits several unique advantages in resolving the issueof small sample data nonlinear data and high-dimensionalpattern recognition [80] SVM builds a hyperplane or aninfinite-dimensional space for classification and regressionThe kernel function allows SVM to deal with the nonlinearclassification problem by attempting to cluster a feature spacebased on the known labels with maximum possible distancebetween the clustersrsquo borders [79] In addition SVM hasbeen widely used in many fields such as text classification[81] image classification [82] biological sequence analysisbiological datamining [83] andhandwriting character recog-nition [84] In recent years SVM has also been applied inthe field of depression discrimination [85ndash87] In the presentstudy GaussianKernel functions have been implemented andevaluated in SVM classification

412 KNN KNN algorithm is a nonparametric supervisedmachine learning method for classification and regression Itwas introduced by Dasarathy [88] in 1991 based on instantor lazy learnings The classifier based on KNN does notrequire a training phase and its computational complexity isproportional to the number of documents in the training setTaken together if the number of documents in the trainingset is 119873 then the time complexity of the KNN classifieris 119874(119899) KNN categorizes the feature spaces into binary ormulticlass clusters by employing a training dataset to furtherclassify the data points according to the closest data pointsto 119870 in the training dataset KNN has been used in medicalinformatics such as the detection of epilepsy [89] stress [90]and depression [85 91]

413 CT CT also known as decision tree is a tree structure-based supervised classification model [92] defined by sepa-rating and partitioning a feature space using multiple rulesand defining a local model into which the feature spaces canbe categorized as binary or multiclass clusters Each of theinternal nodes represents a property each edge represents aresult and each leaf represents a class label Compared to theother classification algorithms the decision tree is the fastestclassification CT has been used in classifying Alzheimerrsquosdisease [93] as well as depression [94]

414 ANN ANN is a classification method that mimics thestructure and function of the biological neural network andconsists of an information processing network with wide par-allel interconnection of simple units This network exhibitslearning and memory ability knowledge generalization andinput information feature extraction ability similar to thatof the human brain [95] Neural networks have been usedto resolve a variety of difficult tasks using common rule-based programming such as computer vision [91] speechrecognition [96] and metal disorders [97 98] ANN is theonly unsupervised machine learning classifier used in thepresent study

42 Classification Result 10-fold cross-validation results ofthe most optimal performance feature combination setsof each classifier and their accuracy in the detection ofdepression are shown below results of resting-state dataneutral audio stimulation data positive audio stimulationdata and negative audio stimulation data are summarizedTables 3 4 5 and 6 respectively

For resting-state EEG data KNN achieved the bestaccuracy of 7683 using feature combination of absolutepower of gamma wave on Fp1 absolute power of theta waveon Fp2 absolute power of beta wave on Fp2 and absolutecenter frequency of beta wave on Fp2 (Table 3)

For EEG data of participants under neutral audio stimu-lation KNN achieved the best accuracy of 7439 using thefeature combination of absolute power of theta on Fp1 centerfrequency of full-band EEG on Fp2 and peak of full-bandEEG on Fp2 (Table 4)

For EEG data of participants under positive audio stimu-lation KNN achieved the best accuracy of 7927 using thefeature combination of absolute power of theta wave on Fp1and absolute power of beta wave on Fp1 (Table 5)

For EEG data of participants under negative audio stim-ulation KNN achieved the best accuracy of 7744 usingfeature combination of absolute power of theta wave on Fp1correlation dimension of full-band EEG on Fp1 absolutecenter frequency of theta wave on Fp2 and absolute powerof gamma wave on Fp2 (Table 6)

The results showed that among all the four classifiers ofSVM KNN CT and ANN KNN performed the best withan average classification accuracy of 7698 (Figure 4) Theabsolute power of theta wave appeared in all the best per-formance feature combination thereby indicating a potentialconnection between theta wave and depressionThe absolutepower of theta wave might be a valid characteristic forpervasive depression discrimination

Complexity 9

Table 3 Classification results in resting state data

Classifier Feature sets Accuracy

SVM Absolute power of theta wave (Fp1) relative power of theta wave (Fp1) relative power of alpha wave (Fp1)absolute center frequency of gamma wave (Fp1) 7256

KNN Absolute power of gamma wave (Fp1) absolute power of theta wave (Fp2) absolute power of beta wave (Fp2)absolute center frequency of beta wave (Fp2) 7683

CT Peak (Fp1) Power-Spectral Entropy of full band EEG (Fp1) absolute power of beta wave (Fp2) 6829

ANN Absolute center frequency of beta wave (Fp1) absolute power of gamma wave (Fp1) Kurtosis of full band EEG(Fp1) absolute power of alpha wave (Fp2) relative center frequency of beta wave (Fp2) 7256

Table 4 Classification results in neutral audio stimulation data


SVM Absolute center frequency of theta wave (Fp1) center frequency of full band EEG (Fp2) peak of full band EEG(Fp2) 7012

KNN Absolute power of theta (Fp1) center frequency of full band EEG (Fp2) peak of full band EEG (Fp2) 7439CT Absolute center frequency of alpha wave (Fp1) Power-Spectral Entropy of alpha wave (Fp1) 6770ANN Relative center frequency of theta wave (Fp2) Hjorth of full band EEG (Fp2) 7378

Table 5 Classification results in positive audio stimulation data

Classifier Feature sets AccuracySVM Absolute power of theta wave (Fp1) Kurtosis of full band EEG (Fp1) 6829KNN Absolute power of theta wave (Fp1) absolute power of beta wave (Fp1) 7927CT Absolute power of gamma wave (Fp1) absolute power of gamma wave (Fp2) 6037

ANNAbsolute power of theta wave (Fp1) power spectral entropy of gamma wave (Fp1) 1198620-complexity of full bandEEG (Fp1) correlation dimension of full band EEG (Fp1) power spectral entropy of theta wave (Fp2)correlation dimension of full band EEG (Fp2)

7439

Table 6 Classification results in negative audio stimulation data

Classifier Feature sets AccuracySVM Hjorth of full band EEG (Fp2) correlation dimension of full band EEG (Fp2) 6707

KNN Absolute power of theta wave (Fp1) correlation dimension of full band EEG (Fp1) absolute center frequency oftheta wave (Fp2) absolute power of gamma wave (Fp2) 7744

CT Absolute power of theta wave (Fp1) power spectral entropy of full band EEG (Fp1) relative power of beta wave(Fp2) peak of full band EEG (Fp2) 7134

ANN Absolute power of theta wave (Fp1) absolute power of beta wave (Fp1) center frequency of full band EEG (Fp1) 7134

5 Conclusion and Future Work

Depression is a major health concern in millions of individu-als Thus diagnosing depression in the early curable stagesis critical for the treatment in order to save the life of apatient However current methods of depression detectionare human-intensive and their results are dependent onthe experience of the doctor Therefore a pervasive andobjective method of diagnosing or even screening wouldbe useful The present study explores a novel method ofdepression detection using pervasive prefrontal-lobe three-electrode EEG system which chooses Fp1 Fp2 and Fpz forelectrode sites according to the international 10-20 system

Several widely employed psychological scales were usedto select the optimal experimental candidates which encom-passed 213 participants (92 depressed patients and 121 normal

controls) Their EEG data of resting state as well as undersound stimulation were recorded The soundtracks wereselected from the IADS-2 database comprising positiveneutral and negative stimuli

The FIR filter combining the Kalman derivation for-mula DWT and an APF were applied on the raw EEGdata to remove the interference from environment ECGEMG and EOG Subsequently 270 linear and nonlinearfeatures were extracted from the preprocessed EEG Thenthe MRMR technique was applied to perform the featureselection Four classification algorithms KNN SVM CTand ANN have been evaluated and compared using a 10-fold cross-validation The results exhibited KNN as the bestperformance classification method in all datasets with thehighest accuracy of 7927 The results also demonstratedthe feature ldquoabsolute power of theta waverdquo in all the best

10 Complexity

6951

7698

6693

7302

SVM KNN CT ANN60006200640066006800700072007400760078008000

()

Figure 4 Average accuracy of classification on selected features

performance features of the four datasets thereby suggestinga robust connection between the power of theta wave anddepression this could be used as a valid characteristic featurein the detection of depression

The current study postulated that a novel and pervasivesystem for screening depression is feasible With a carefullydesigned model the pervasive system could reach accuracysimilar to the current scale-based screening method forinstance the accuracy of BDI is reported to be 79ndash86 indifferent studies [99ndash101]

EEG and depression have been under intensive focus ofresearch In comparison to the study by Knott et al whocollected EEG recordings from 21 scalp sites and conductedunivariate analyses for group comparisons and correctlyclassified 913 of the patients and controls [102] the currentpervasive three-electrode EEG acquisition system can bemeasured easily and is rather suitable for the personaluse of patients Moreover we used a less number of elec-trodes thereby reducing the amount of data considerablyHealey and Picard attempted to extract feature patternsfrom physiological signals for emotional recognition withan accuracy of 618ndash784 [103] Kim et al developed ashort-term monitoring emotion recognition system basedon multiuser physiological signals to extract features andidentify three emotions using SVM The final recognitionaccuracywas 75 [104] Compared to these studies our resulthas a higher accuracy with faster data processing efficiencyHenriques collected the resting-state EEG (with closed eyes)from 5 depressed patients and 13 normal individuals Theresults showed that the activity in the left hemisphere indepressed patients is significantly lower than that in a normalperson [105] Omelrsquochenko and Zaika collected EEG datafrom 53 depressed patients and 86 normal individuals anddemonstrated that patients with depression have a higherdelta and theta energy than normal but a lower alpha andbeta energy [106] Fingelkurts et al researched the shockcomponents of resting-state EEG from 12 depressed patientsand 10 normal persons and found that the brain activates wereaffected by depression throughout the cerebral cortex [107]Compared to these studies the current experiment presentedmore reliable data to ensure the reliability of the experimentalresults Taken together our data model based on featureextraction and feature selection reduced the amount of data

to be processed with a faster data processing efficiencyIn addition the pervasive three-electrode EEG acquisitionsystemwas easier tomeasure as well as ensuring the accuracyof data

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by the National Basic ResearchProgram of China (973 Program) (no 2014CB744600) theNational Natural Science Foundation of China (Grants nos61210010 and 61632014) and the Program of InternationalSampT Cooperation of MOST (no 2013DFA11140)

References

[1] M E P SeligmanHelplessness onDepression Development andDeath WH FreemanTimes BooksHenry Holt amp Co 1975

[2] ldquoWorld Health Organization Fact sheet on depression [EBOL][2017-02]rdquo httpwwwwhointmediacentrefactsheetsfs369en

[3] World Health Organization The World Health Report 2001Mental health new understanding new hope World HealthOrganization 2001

[4] World Health Organization Sixty-Fifth World Health Assembly(2012) [EBOL] 2012 httpwwwwhointmediacentreevents2012wha65en

[5] American Psychiatric Association Diagnostic and StatisticalManual of Mental Disorders vol 1 American Psychiatric Asso-ciation Arlington Va USA 4th edition 2000

[6] V C Pangman J Sloan and L Guse ldquoAn examination of psy-chometric properties of the mini-mental state examination andthe standardized mini-mental state examination implicationsfor clinical practicerdquoApplied Nursing Research vol 13 no 4 pp209ndash213 2000

[7] A BeckT R SteerA andK BrownGBeck depression inventory1996

[8] M Hamilton ldquoA rating scale for depressionrdquo Journal of Neurol-ogy Neurosurgery amp Psychiatry vol 23 pp 56ndash62 1960

[9] B Lindsley D Emotions and the electroencephalogram 1950[10] B Wilder ldquoEEG in clinical practicerdquo Electroencephalography

and Clinical Neurophysiology vol 56 no 5 p 536 1983[11] K G Jordan ldquoContinuous EEG and evoked potential monitor-

ing in the neuroscience intensive care unitrdquo Journal of ClinicalNeurophysiology vol 10 no 4 pp 445ndash475 1993

[12] J C Woestenburg M N Verbaten and J L Slangen ldquoTheremoval of the eye-movement artifact from the EEG by regres-sion analysis in the frequency domainrdquo Biological Psychologyvol 16 no 1-2 pp 127ndash147 1983

[13] W Klimesch ldquoEEG alpha and theta oscillations reflect cogni-tive and memory performance a review and analysisrdquo BrainResearch Reviews vol 29 no 2-3 pp 169ndash195 1999

[14] H W Cole and W J Ray ldquoEEG correlates of emotionaltasks related to attentional demandsrdquo International Journal ofPsychophysiology vol 3 no 1 pp 33ndash41 1985

Complexity 11

[15] W Klimesch M Doppelmayr H Russegger T Pachinger andJ Schwaiger ldquoInduced alpha band power changes in the humanEEG and attentionrdquoNeuroscience Letters vol 244 no 2 pp 73ndash76 1998

[16] R Srinivasan S Thorpe S Deng T Lappas and M DrsquoZmuraldquoDecoding attentional orientation from eeg spectrardquo LectureNotes in Computer Science (including subseries Lecture Notesin Artificial Intelligence and Lecture Notes in Bioinformatics)Preface vol 5610 no 1 pp 176ndash183 2009

[17] E A Curran and M J Stokes ldquoLearning to control brain activ-ity A review of the production and control of EEG componentsfor driving brain-computer interface (BCI) systemsrdquo Brain andCognition vol 51 no 3 pp 326ndash336 2003

[18] A S Gevins G M Zeitlin C D Yingling et al ldquoEEG patternsduring rsquocognitiversquo tasks I Methodology and analysis of complexbehaviorsrdquo Electroencephalography and Clinical Neurophysiol-ogy vol 47 no 6 pp 693ndash703 1979

[19] H Haken Principles of brain functioning a synergetic approachto brain activity behavior and cognition Springer Science ampBusiness Media 2013

[20] F Fan Y Li Y Qiu and Y Zh ldquoUse of ANN and ComplexityMeasures in Cognitive EEG Discriminationrdquo in Proceedings ofthe 2005 IEEE Engineering in Medicine and Biology 27th AnnualConference pp 4638ndash4641 Shanghai China 2005

[21] H Merica R Blois and J-M Gaillard ldquoSpectral characteristicsof sleep EEG in chronic insomniardquo European Journal of Neuro-science vol 10 no 5 pp 1826ndash1834 1998

[22] R R Rosa and M H Bonnet ldquoReported chronic insomnia isindependent of poor sleep asmeasured by electroencephalogra-phyrdquo Psychosomatic Medicine vol 62 no 4 pp 474ndash482 2000

[23] H Merica and J-M Gaillard ldquoThe EEG of the sleep onsetperiod in insomnia A discriminant analysisrdquo Physiology ampBehavior vol 52 no 2 pp 199ndash204 1992

[24] H Adeli S Ghosh-Dastidar and N Dadmehr ldquoA wavelet-chaos methodology for analysis of EEGs and EEG subbands todetect seizure and epilepsyrdquo IEEE Transactions on BiomedicalEngineering vol 54 no 2 pp 205ndash211 2007

[25] F Mormann K Lehnertz P David and C E Elger ldquoMeanphase coherence as a measure for phase synchronization andits application to the EEG of epilepsy patientsrdquo Physica DNonlinear Phenomena vol 144 no 3 pp 358ndash369 2000

[26] K Lehneltz G R J Arnhold and T Kreuz ldquoNonlinear EEGAnalysis in Epilepsy Its Possible Use for Interictal FocusLocalization Seizure Anticipation and Preventionrdquo Joumal ofClinical Neurophysiology vol 18 no 3 pp 209ndash222

[27] R Williams L I Karacan and C J Hursch Electroencephalog-raphy (EEG) of human sleep clinical applications John Wiley ampSons 1974

[28] D K Hannesdottir J Doxie M A Bell T H Ollendick andC D Wolfe ldquoA longitudinal study of emotion regulation andanxiety in middle childhood Associations with frontal EEGasymmetry in early childhoodrdquo Developmental Psychobiologyvol 52 no 2 pp 197ndash204 2010

[29] O Siciliani M Schiavon and M Tansella ldquoAnxiety and EEGalpha activity in neurotic patientsrdquoActa Psychiatrica Scandinav-ica vol 52 no 2 pp 116ndash131 1975

[30] J Avram F R Baltes M Miclea and A C Miu ldquoFrontalEEG activation asymmetry reflects cognitive biases in anxietyEvidence from an emotional face Stroop taskrdquo Applied Psy-chophysiology and Biofeedback vol 35 no 4 pp 285ndash292 2010

[31] B A Clementz S R Sponheim W G Iacono and M BeiserldquoResting EEG in first-episode schizophrenia patients bipolarpsychosis patients and their first-degree relativesrdquo Psychophys-iology vol 31 no 5 pp 486ndash494 1994

[32] R Diaz-guerrero J S Gottlieb and J R Knott ldquoThe sleep ofpatients with manic-depressive psychosis depressive type anelectroencephalographic studyrdquo Psychosomatic Medicine vol 8no 6 pp 399ndash404 1946

[33] P A Davis and H Davis ldquoThe electroencephalograms ofpsychotic patientsrdquoThe American Journal of Psychiatry vol 95no 5 pp 1007ndash1025 1939

[34] S R Sponheim B A Clementz W G Iacono and M BeiserldquoClinical and biological concomitants of resting state EEGpower abnormalities in schizophreniardquo Biological Psychiatryvol 48 no 11 pp 1088ndash1097 2000

[35] J Bures O Buresova and J KıivanekThemechanism and appli-cations of Leaorsquos spreading depression of electroencephalographicactivity Academic Press 1974

[36] G E Bruder R Fong C E Tenke et al ldquoRegional brainasymmetries in major depression with or without an anxietydisorder A quantitative electroencephalographic studyrdquoBiolog-ical Psychiatry vol 41 no 9 pp 939ndash948 1997

[37] S A Reid L M Duke and J J B Allen ldquoResting frontal elec-troencephalographic asymmetry in depression Inconsistenciessuggest the need to identify mediating factorsrdquo Psychophysiol-ogy vol 35 no 4 pp 389ndash404 1998

[38] RThibodeau R S Jorgensen and S Kim ldquoDepression anxietyand resting frontal EEG asymmetry a meta-analytic reviewrdquoJournal of Abnormal Psychology vol 115 no 4 pp 715ndash7292006

[39] N Liang P Saratchandran G Huang and N SundararajanldquoClassification of mental tasks from EEG signals using extremelearning machinerdquo International Journal of Neural Systems vol16 no 1 pp 29ndash38 2006

[40] C W Anderson E A Stolz and S Shamsunder ldquoMulti-variate autoregressive models for classification of spontaneouselectroencephalographic signals during mental tasksrdquo IEEETransactions on Biomedical Engineering vol 45 no 3 pp 277ndash286 1998

[41] J D R Millan J Mourino M Franze et al ldquoA local neuralclassifier for the recognition of EEG patterns associated tomental tasksrdquo IEEE Transactions on Neural Networks andLearning Systems vol 13 no 3 pp 678ndash686 2002

[42] T Harmony T Fernandez J Silva et al ldquoEEG delta activityAn indicator of attention to internal processing during perfor-mance of mental tasksrdquo International Journal of Psychophysiol-ogy vol 24 no 1-2 pp 161ndash171 1996

[43] ldquoElectrical Geodesics Incrdquo httpswwwegicom[44] ldquoBrain Products GmbHrdquo httpwwwbrainproductscom[45] ldquoUAIS Lab of Lanzhou Universityrdquo httpwwwtheuaisorg[46] H H Jasper ldquoThe ten twenty electrode system of the interna-

tional federationrdquo Electroencephalography and Clinical Neurophsiology vol 10 pp 371ndash375 1958

[47] G E Chatrian E Lettich and P L Nelson ldquoTen percentelectrode system for topographic studies of spontaneous andevoked EEG activitiesrdquo American Journal of EEG Technologyvol 25 no 2 pp 83ndash92 1985

[48] S Sanei and J A Chambers EEG signal processing John Wileyamp Sons 2013

[49] H Jasper ldquoThe ten twenty electrode system of the introductionfederationrdquo Electroencephalography and Clinical Neurophysiol-ogy vol 10 pp 371ndash375 1958

12 Complexity

[50] W J H Nauta ldquoThe problem of the frontal lobe A reinterpreta-tionrdquo Journal of Psychiatric Research vol 8 no 3-4 pp 167ndash1871971

[51] E T Rolls ldquoOn the brain and emotionrdquo Behavioral and BrainSciences vol 23 no 2 pp 219ndash228 2000

[52] E Harmon-Jones ldquoContributions from research on angerand cognitive dissonance to understanding the motivationalfunctions of asymmetrical frontal brain activityrdquo BiologicalPsychology vol 67 no 1-2 pp 51ndash76 2004

[53] Lanzhou University ldquoA portable EEG acquisition systemrdquoChina CN2015206281526 2016-02-17

[54] J-M Azorin P Benhaım T Hasbroucq and C-A PossamaıldquoStimulus preprocessing and response selection in depressionA reaction time studyrdquo Acta Psychologica vol 89 no 2 pp 95ndash100 1995

[55] W O Tatum B A Dworetzky and D L Schomer ldquoArtifact andrecording concepts in EEGrdquo Journal of Clinical Neurophysiologyvol 28 no 3 pp 252ndash263 2011

[56] A T Beck Depression Clinical experimental and theoreticalaspects University of Pennsylvania Press 1967

[57] J Epstein H Pan J H Kocsis et al ldquoLack of ventral striatalresponse to positive stimuli in depressed versus normal sub-jectsrdquo The American Journal of Psychiatry vol 163 no 10 pp1784ndash1790 2006

[58] L M Bylsma B H Morris and J Rottenberg ldquoA meta-analysisof emotional reactivity in major depressive disorderrdquo ClinicalPsychology Review vol 28 no 4 pp 676ndash691 2008

[59] E-M Yong W-Q Qian and K-F He ldquoPenetration abilityanalysis for glide reentry trajectory based on radar trackingrdquoYuhang XuebaoJournal of Astronautics vol 33 no 10 pp 1370ndash1376 2012

[60] W Ende Z Feng X Yanghui T Xinxin and Z Dan ldquoGen-erating Method of Adaptive Linear Signal Based on KalmanPredictionrdquoModern Defense Technology vol 40 no 2 pp 138ndash142 2012

[61] Z Peng ldquoAero engine Fault Diagnostics Based on KalmanFilterrdquoNanjingUniversity of Aeronautics andAstronautics 2008

[62] Z Xiao W Ming and L Xiaoming ldquopplication of extendedKalmanfilter algorithm in intelligent robot soccer competitionrdquoElectrical and Mechanical Engineering vol 29 pp 334ndash3382012

[63] H Singh and J Singh ldquoA review on electrooculographyrdquoInternational Journal of Advanced Engineering Technology vol3 pp 115ndash122

[64] L Yang Study of Ocular Artifacts Removal Based on WaveletTransform and Kalman filter Lanzhou University 2016

[65] S Tong A Bezerianos J Paul Y Zhu and N ThakorldquoRemoval of ECG interference from the EEG recordings insmall animals using independent component analysisrdquo Journalof Neuroscience Methods vol 108 no 1 pp 11ndash17 2001

[66] AGevins SWDu andH LeongAdaptive interference cancelerfor EEG movement and eye artifacts US Patent 5513649[P]1996-5-7

[67] J P M Pijn Quantitative evaluation of EEG signals in epilepsynonlinear association time delays and nonlinear dynamics [PhDthesis] University of Amsterdam 1990

[68] J P M Pijn D N Veils M J Van Der Heyden J DeGoedeC W M Van Veelen and F H Lopes Da Silva ldquoNonlineardynamics of epileptic seizures on basis of intracranial EEGrecordingsrdquo Brain Topography vol 9 no 4 pp 249ndash270 1997

[69] S A R B Rombouts R W M Keunen and C J StamldquoInvestigation of nonlinear structure in multichannel EEGrdquoPhysics Letters A vol 202 no 5-6 pp 352ndash358 1995

[70] C J Stam T C A M Van Woerkom and R W MKeunen ldquoNon-linear analysis of the electroencephalogram inCreutzfeldt-Jakob diseaserdquo Biological Cybernetics vol 77 no 4pp 247ndash256 1997

[71] U R Acharya O Faust N Kannathal T Chua and SLaxminarayan ldquoNon-linear analysis of EEG signals at varioussleep stagesrdquo Computer Methods and Programs in Biomedicinevol 80 no 1 pp 37ndash45 2005

[72] B Hjorth ldquoEEG analysis based on time domain propertiesrdquoElectroencephalography and Clinical Neurophysiology vol 29no 3 pp 306ndash310 1970

[73] E H Shen Z J Cai and F J Gu ldquoMathematical foundation of anew complexity measurerdquo Applied Mathematics andMechanicsYingyong Shuxue He Lixue vol 26 no 9 pp 1083ndash1090 2005

[74] A Lempel and J Ziv ldquoOn the Complexity of Finite SequencesrdquoIEEE Transactions on Information Theory vol 22 no 1 pp 75ndash81 1976

[75] C E Shannon ldquoAmathematical theory of communicationrdquo BellLabs Technical Journal vol 27 pp 379ndash423 623ndash656 1948

[76] J Bruhn L E Lehmann H Ropcke T W Bouillon and AHoeft ldquoShannon entropy applied to the measurement of theelectroencephalographic effects of desfluranerdquo Anesthesiologyvol 95 no 1 pp 30ndash35 2001

[77] C E Shannon ldquoA mathematical theory of communicationrdquoAcm Sigmobile Mobile Computing amp Communications Reviewvol 5 no 1 pp 3ndash55 1948

[78] H Peng F Long and C Ding ldquoFeature selection basedon mutual information criteria of max-dependency max-relevance and min-redundancyrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 27 no 8 pp 1226ndash12382005

[79] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[80] M Hearst S Dumais E Osman J Platt and B ScholkopfldquoSupport vector machinesrdquo IEEE Intelligent Systems and TheirApplications vol 13 no 4 pp 18ndash28 1998

[81] T Joachims ldquoText categorizationwith support vectormachineslearning with many relevant featuresrdquo in Machine LearningECML-98 vol 1398 pp 137ndash142 Springer Berlin Germany1998

[82] B Gaonkar andCDavatzikos ldquoAnalytic estimation of statisticalsignificance maps for support vector machine based multi-variate image analysis and classificationrdquo NeuroImage vol 78pp 270ndash283 2013

[83] R Cuingnet C Rosso M Chupin et al ldquoSpatial regularizationof SVM for the detection of diffusion alterations associated withstroke outcomerdquoMedical Image Analysis vol 15 no 5 pp 729ndash737 2011

[84] C Bahlmann B Haasdonk and H Burkhardt ldquoOnline hand-writing recognition with support vector machines - A kernelapproachrdquo in Proceedings of the 8th International Workshop onFrontiers in Handwriting Recognition IWFHR 2002 pp 49ndash54Canada August 2002

[85] O Faust P C A Ang S D Puthankattil and P K JosephldquoDepression diagnosis support system based on eeg signalentropiesrdquo Journal of Mechanics in Medicine and Biology vol14 no 3 Article ID 1450035 2014

Complexity 13

[86] B Hosseinifard M H Moradi and R Rostami ldquoClassify-ing depression patients and normal subjects using machinelearning techniquesrdquo in Proceedings of the 2011 19th IranianConference on Electrical Engineering ICEE 2011 irn May 2011

[87] I Kalatzis N Piliouras E Ventouras C Papageorgiou ARabavilas and D Cavouras ldquoComparative evaluation of prob-abilistic neural network versus support vector machines clas-sifiers in discriminating ERP signals of depressive patientsfrom healthy controlsrdquo in Proceedings of the 3rd InternationalSymposium on Image and Signal Processing and Analysis 2003ISPA 2003 pp 981ndash985 Rome Italy

[88] B V Dasarathy Nearest neighbor (NN) norms NN patternclassification techniques 1991

[89] 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

[90] J-S Wang C-W Lin and Y-T C Yang ldquoA k-nearest-neighborclassifier with heart rate variability feature-based transforma-tion algorithm for driving stress recognitionrdquo Neurocomputingvol 116 pp 136ndash143 2013

[91] H A Rowley S Baluja and T Kanade ldquoNeural network-based face detectionrdquo IEEETransactions on PatternAnalysis andMachine Intelligence vol 20 no 1 pp 23ndash38 1998

[92] M Hansen R Dubayah and R Defries ldquoClassification treesAn alternative to traditional land cover classifiersrdquo InternationalJournal of Remote Sensing vol 17 no 5 pp 1075ndash1081 1996

[93] C Lehmann T Koenig V Jelic et al ldquoApplication and compar-ison of classification algorithms for recognition of Alzheimerrsquosdisease in electrical brain activity (EEG)rdquo Journal of Neuro-science Methods vol 161 no 2 pp 342ndash350 2007

[94] U R Acharya V K Sudarshan H Adeli et al ldquoA noveldepression diagnosis index using nonlinear features in EEGsignalsrdquo European Neurology vol 74 no 1-2 pp 79ndash83 2016

[95] S Dreiseitl and L Ohno-Machado ldquoLogistic regression andartificial neural network classification models a methodologyreviewrdquo Journal of Biomedical Informatics vol 35 no 5-6 pp352ndash359 2002

[96] G Hinton L Deng D Yu et al ldquoDeep neural networks foracoustic modeling in speech recognition the shared views offour research groupsrdquo IEEE Signal Processing Magazine vol 29no 6 pp 82ndash97 2012

[97] S D Puthankattil and P K Joseph ldquoClassification of eeg signalsin normal and depression conditions by ann using rwe andsignal entropyrdquo Journal of Mechanics in Medicine and Biologyvol 12 no 4 Article ID 1240019 2012

[98] Y Li and F Fan ldquoClassification of Schizophrenia and depressionbyEEGwithANNsrdquo inProceedings of the 2005 IEEEEngineeringinMedicine andBiology 27thAnnual Conference pp 2679ndash2682Engineering inMedicine and Biology Society Shanghai ChinaJanuary 2006

[99] A FischerM Fischer R ANicholls et al ldquoDiagnostic accuracyfor major depression in multiple sclerosis using self-reportquestionnairesrdquo Brain and Behavior vol 5 no 9 Article IDe00365 2015

[100] I M Cameron A Cardy J R Crawford et al ldquoMeasuringdepression severity in general practice Discriminatory perfor-mance of the PHQ-9 HADS-D and BDI-IIrdquo British Journal ofGeneral Practice vol 61 no 588 pp e419ndashe426 2011

[101] C D Silberman J Laks C F Capitao C S RodriguesI Moreira and E Engelhardt ldquoRecognizing depression in

patients with Parkinsonrsquos disease Accuracy and specificity oftwo depression rating scalerdquoArquivos de Neuro-Psiquiatria vol64 no 2 B pp 407ndash411 2006

[102] V Knott C Mahoney S Kennedy and K Evans ldquoEEG powerfrequency asymmetry and coherence in male depressionrdquoPsychiatry Research Neuroimaging vol 106 no 2 pp 123ndash1402001

[103] J Healey and R Picard ldquoDigital processing of affective signalsrdquoin Proceedings of the 1998 23rd IEEE International Conference onAcoustics Speech and Signal Processing ICASSP 1998 pp 3749ndash3752 USA May 1998

[104] K H Kim S W Bang and S R Kim ldquoEmotion recognitionsystem using short-term monitoring of physiological signalsrdquoMedical amp Biological Engineering amp Computing vol 42 no 3pp 419ndash427 2004

[105] J B Henriques and R J Davidson ldquoLeft Frontal Hypoactivationin Depressionrdquo Journal of Abnormal Psychology vol 100 no 4pp 535ndash545 1991

[106] V P Omelrsquochenko and V G Zaika ldquoChanges in the EEGrhythms in endogenous depressive disorders and the effect ofpharmacotherapyrdquo Fiziologiia cheloveka vol 28 no 3 pp 30ndash36 2002

[107] A A Fingelkurts A A Fingelkurts H Rytsala K Suominen EIsometsa and S Kahkonen ldquoComposition of brain oscillationsin ongoing EEG during major depression disorderrdquo Neuro-science Research vol 56 no 2 pp 133ndash144 2006

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of


Mathematical Problems in Engineering

Applied MathematicsJournal of


Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of


Mathematical PhysicsAdvances in

Complex AnalysisJournal of


OptimizationJournal of



Engineering Mathematics

International Journal of


Operations ResearchAdvances in

Journal of


Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences


Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018


Decision SciencesAdvances in


AnalysisInternational Journal of


Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

2 Complexity

psychiatrist The current international standard mostly usedis ldquoIn Diagnostic and Statistical Manual of Mental Disorders(Fourth Edition)rdquo (DSM-IV) [5] and the clinical test Mini-Mental State Examination (MMSE) is commonly applied[6] Other conventional psychometric questionnaires such asBeck depression inventory (BDI) [7] and Hamilton Depres-sion Rating Scale (HDRS) [8] are also used as screening toolsrather than as the instrument for the diagnosis of depression

The current methods of depression detection are human-intensive and the results are dependent on the doctorrsquosexperience Furthermore depressed individuals are less likelyto seek help due to fear of stigma and the nature of thedisorder As a result a large number of depressed patients notdiagnosed accurately do not receive optimal treatment andadequate recovery period Therefore finding convenient andeffectivemethods for the detection of depression is an emerg-ing topic for research With the latest advances in the sensorand mobile technology the exploration using physiologicaldata for the diagnosis of mental disorder opens a new avenuefor an objective and accurate tool for depression detectionAmong all kinds of physiological data electroencephalogram(EEG) reflects emotional human brain activity in real time[9]

The EEG signal is a recording of the spontaneousrhythmic electrical activity of brain neurons from the scalpsurface Since the earliest discovery from the rabbit andmonkey brain and the first recording of the human EEGsignal by German psychiatrist Hans Berger in 1926 studieson the analytical method of EEG and the interpretation of theassociation between the brain function and mental disordershave been continued for over a century [10] Neurosciencepsychology and cognitive science research showed that amajority of the psychological activities and cognitive behav-ior could be indicated by EEG [11ndash13] The EEG signal isclosely related to the brain activities and emotional statesand it could reflect the emotional transformation in real timeCole and Ray [14] found that the EEG signal collected fromthe parietal lobe of brain is associated with the cognitive tasksand emotional states Klimesch et al found that the alphawaves with low frequency could reflect some of the features ofattention such as vigilance and expectations [15] Srinivasanet al demonstrated that the frequency domain features ofEEG could be used to predict the level of attention [16]Therefore the EEG signal is critical for understanding theprocessing of human brain information and emotional statetransformation

The studies on EEG could be used to understand themechanism underlying brain activity human cognitive pro-cess and diagnosis of brain disease as well as the field ofthe Brain Computer Interface (BCI) which has attractedmuch attention in recent years [17] Compared to Com-puted Tomography (CT) and functionalMagnetic ResonanceImaging (fMRI) EEG has a higher time resolution a lowermaintenance cost and a simpler operation method Thusas an objective physiological method to obtain data EEGwas proposed as a nonintrusive approach to study cogni-tive behavior [18ndash20] and other illness symptoms such asinsomnia [21ndash23] epilepsy [24ndash26] and sleep disorder [27]EEG has also been used in the diagnosis of mental disorders

such as anxiety [28ndash30] psychosis [31ndash34] and depression[35ndash38] In addition depression as a mental disorder withclinical manifestations such as significant depression andslow thinking is always accompanied by abnormal brainactivity and obvious emotional alternation Therefore as amethod tracking the brain functions EEG can detect theseabnormal activities

The frequency of the EEG signal can be divided into 5wave-bands delta wave (lt4Hz) which normally appears inan adultrsquos slow-wave sleep theta wave (4ndash8Hz) which isusually found when someone is sleepy alpha wave (8ndash14Hz)which is normally detected when someone is relaxed betawave (14ndash30Hz) which commonly appears when someone isactively thinking and gamma wave (30ndash50Hz) which couldappear during meditationThe EEG signals undergo changesin the amplitude as well as frequency while different mentaltasks are performed [39ndash42]

Presently for research purposes the most commonlyused are 128-electrode and 256-electrode EEG systems [4344] which are specifically designed for research purposesThe operation of the instruments was not only difficult toinitiate but also it required technicians to apply conductivegel to each electrode on the participantrsquos head before each useThe preparation process alone takes 30 minutes on averageIn addition these EEG systems are expensive Overall thesesystems are not practical for pervasive depression detec-tion

In the present study the pervasive three-electrode EEGacquisition system developed independently by the Ubiq-uitous Awareness and Intelligent Solutions Lab (UAIS) ofLanzhou University [45] was employed to construct adatabase containing both depressed patients and normalcontrols Thus the use of the latest data processing tech-nique and machine learning to explore a pervasive EEG-based depression detection system has been the focus ofinvestigation In order to support this research

(1) A pervasive three-electrode EEG acquisition systemhas been introduced (Section 21)

(2) A psychophysiological experiment has been con-ducted in which EEG of 213 participants has beenrecorded These physiological data provided a com-prehensive database for further analysis construc-tion and evaluation of a pervasive EEG-based depres-sion detection system (Sections 22 and 23)

(3) Several EEG preprocessing steps and methods wereapplied on the raw EEG data (Section 31)

(4) 270 features were identified and extracted from therecoded database By employing a feature selectiontechnique an optimum feature matrix was con-structed for the depression classification process (Sec-tion 32)

(5) Four classification algorithms including K-NearestNeighbor (KNN) Support Vector Machine (SVM)Classification Tree (CT) and Artificial Neural Net-work (ANN) have been evaluated and comparedusing a 10-fold cross-validation (Section 4)

Complexity 3

C4CzC3

F3F7

Fp1 Fp2

Fz F4 F8

T3

T5 T6

O2O1

T4

P4PzP3

A2A1

Nasion

Inion










4 Complexity

EEG signal

Ground

FP2 FPZ FP1



















Complexity 5


3 Data Processing






4119898119873 minus 1)]

sdot 119877119873 (119899) (1)




119910 (119899) = 119909 (119899) + 119865 (119903) (2)









6 Complexity







119883 (119896) = 1119873119873minus1sum119899=0

119909 (119899) 119890minus119895(2119896120587119899119873) 119896 = 0 1 119873 minus 1 (3)


119872 = 1119873119873minus1sum119896=0

|119883 (119896)|2 (4)


119884 (119896) = 119883(119896) |119883 (119896)|2 gt 1198720 |119883 (119896)|2 le 119872 (5)


119896=0

119884 (119896) 119890119895(2119896120587119899119873) 119899 = 0 1 119873 minus 1 (6)


1198601 =119873minus1sum119899=0

1003816100381610038161003816119909 (119899) minus 119910 (119899)10038161003816100381610038162

1198600 =119873minus1sum119899=0

|119909 (119899)|2

1198620 = 11986011198600

(7)



lim120576rarr0

lim119899rarrinfin



Complexity 7


119867(119883) = sum119909isin120594

119901 (119909) log119901 (119909) (9)



CD = lim119903rarr0

( ln119862 (119903)ln 119903 ) (10)







Linear features





max119863 (119878 119888) = 1|119878| sum119891119894isin119878

Φ(119891119894 119888) (11)



min119877 (119878) = 1|119878|2 sum119891119894 119891119895isin119878

Φ(119891119894 119891119895) (12)


8 Complexity


max119891119895notin119878119898minus1

[[Φ (119891119894 119888) minus 1


Φ(119891119894 119891119895)]] (13)


4 Classification













Complexity 9














7439











10 Complexity

6951

7698

6693

7302

SVM KNN CT ANN60006200640066006800700072007400760078008000

()








Acknowledgments


References















Complexity 11


































12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








Complexity 3

C4CzC3

F3F7

Fp1 Fp2

Fz F4 F8

T3

T5 T6

O2O1

T4

P4PzP3

A2A1

Nasion

Inion










4 Complexity

EEG signal

Ground

FP2 FPZ FP1



















Complexity 5


3 Data Processing






4119898119873 minus 1)]

sdot 119877119873 (119899) (1)




119910 (119899) = 119909 (119899) + 119865 (119903) (2)









6 Complexity







119883 (119896) = 1119873119873minus1sum119899=0

119909 (119899) 119890minus119895(2119896120587119899119873) 119896 = 0 1 119873 minus 1 (3)


119872 = 1119873119873minus1sum119896=0

|119883 (119896)|2 (4)


119884 (119896) = 119883(119896) |119883 (119896)|2 gt 1198720 |119883 (119896)|2 le 119872 (5)


119896=0

119884 (119896) 119890119895(2119896120587119899119873) 119899 = 0 1 119873 minus 1 (6)


1198601 =119873minus1sum119899=0

1003816100381610038161003816119909 (119899) minus 119910 (119899)10038161003816100381610038162

1198600 =119873minus1sum119899=0

|119909 (119899)|2

1198620 = 11986011198600

(7)



lim120576rarr0

lim119899rarrinfin



Complexity 7


119867(119883) = sum119909isin120594

119901 (119909) log119901 (119909) (9)



CD = lim119903rarr0

( ln119862 (119903)ln 119903 ) (10)







Linear features





max119863 (119878 119888) = 1|119878| sum119891119894isin119878

Φ(119891119894 119888) (11)



min119877 (119878) = 1|119878|2 sum119891119894 119891119895isin119878

Φ(119891119894 119891119895) (12)


8 Complexity


max119891119895notin119878119898minus1

[[Φ (119891119894 119888) minus 1


Φ(119891119894 119891119895)]] (13)


4 Classification













Complexity 9














7439











10 Complexity

6951

7698

6693

7302

SVM KNN CT ANN60006200640066006800700072007400760078008000

()








Acknowledgments


References















Complexity 11


































12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








4 Complexity

EEG signal

Ground

FP2 FPZ FP1



















Complexity 5


3 Data Processing






4119898119873 minus 1)]

sdot 119877119873 (119899) (1)




119910 (119899) = 119909 (119899) + 119865 (119903) (2)









6 Complexity







119883 (119896) = 1119873119873minus1sum119899=0

119909 (119899) 119890minus119895(2119896120587119899119873) 119896 = 0 1 119873 minus 1 (3)


119872 = 1119873119873minus1sum119896=0

|119883 (119896)|2 (4)


119884 (119896) = 119883(119896) |119883 (119896)|2 gt 1198720 |119883 (119896)|2 le 119872 (5)


119896=0

119884 (119896) 119890119895(2119896120587119899119873) 119899 = 0 1 119873 minus 1 (6)


1198601 =119873minus1sum119899=0

1003816100381610038161003816119909 (119899) minus 119910 (119899)10038161003816100381610038162

1198600 =119873minus1sum119899=0

|119909 (119899)|2

1198620 = 11986011198600

(7)



lim120576rarr0

lim119899rarrinfin



Complexity 7


119867(119883) = sum119909isin120594

119901 (119909) log119901 (119909) (9)



CD = lim119903rarr0

( ln119862 (119903)ln 119903 ) (10)







Linear features





max119863 (119878 119888) = 1|119878| sum119891119894isin119878

Φ(119891119894 119888) (11)



min119877 (119878) = 1|119878|2 sum119891119894 119891119895isin119878

Φ(119891119894 119891119895) (12)


8 Complexity


max119891119895notin119878119898minus1

[[Φ (119891119894 119888) minus 1


Φ(119891119894 119891119895)]] (13)


4 Classification













Complexity 9














7439











10 Complexity

6951

7698

6693

7302

SVM KNN CT ANN60006200640066006800700072007400760078008000

()








Acknowledgments


References















Complexity 11


































12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








Complexity 5


3 Data Processing






4119898119873 minus 1)]

sdot 119877119873 (119899) (1)




119910 (119899) = 119909 (119899) + 119865 (119903) (2)









6 Complexity







119883 (119896) = 1119873119873minus1sum119899=0

119909 (119899) 119890minus119895(2119896120587119899119873) 119896 = 0 1 119873 minus 1 (3)


119872 = 1119873119873minus1sum119896=0

|119883 (119896)|2 (4)


119884 (119896) = 119883(119896) |119883 (119896)|2 gt 1198720 |119883 (119896)|2 le 119872 (5)


119896=0

119884 (119896) 119890119895(2119896120587119899119873) 119899 = 0 1 119873 minus 1 (6)


1198601 =119873minus1sum119899=0

1003816100381610038161003816119909 (119899) minus 119910 (119899)10038161003816100381610038162

1198600 =119873minus1sum119899=0

|119909 (119899)|2

1198620 = 11986011198600

(7)



lim120576rarr0

lim119899rarrinfin



Complexity 7


119867(119883) = sum119909isin120594

119901 (119909) log119901 (119909) (9)



CD = lim119903rarr0

( ln119862 (119903)ln 119903 ) (10)







Linear features





max119863 (119878 119888) = 1|119878| sum119891119894isin119878

Φ(119891119894 119888) (11)



min119877 (119878) = 1|119878|2 sum119891119894 119891119895isin119878

Φ(119891119894 119891119895) (12)


8 Complexity


max119891119895notin119878119898minus1

[[Φ (119891119894 119888) minus 1


Φ(119891119894 119891119895)]] (13)


4 Classification













Complexity 9














7439











10 Complexity

6951

7698

6693

7302

SVM KNN CT ANN60006200640066006800700072007400760078008000

()








Acknowledgments


References















Complexity 11


































12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








6 Complexity







119883 (119896) = 1119873119873minus1sum119899=0

119909 (119899) 119890minus119895(2119896120587119899119873) 119896 = 0 1 119873 minus 1 (3)


119872 = 1119873119873minus1sum119896=0

|119883 (119896)|2 (4)


119884 (119896) = 119883(119896) |119883 (119896)|2 gt 1198720 |119883 (119896)|2 le 119872 (5)


119896=0

119884 (119896) 119890119895(2119896120587119899119873) 119899 = 0 1 119873 minus 1 (6)


1198601 =119873minus1sum119899=0

1003816100381610038161003816119909 (119899) minus 119910 (119899)10038161003816100381610038162

1198600 =119873minus1sum119899=0

|119909 (119899)|2

1198620 = 11986011198600

(7)



lim120576rarr0

lim119899rarrinfin



Complexity 7


119867(119883) = sum119909isin120594

119901 (119909) log119901 (119909) (9)



CD = lim119903rarr0

( ln119862 (119903)ln 119903 ) (10)







Linear features





max119863 (119878 119888) = 1|119878| sum119891119894isin119878

Φ(119891119894 119888) (11)



min119877 (119878) = 1|119878|2 sum119891119894 119891119895isin119878

Φ(119891119894 119891119895) (12)


8 Complexity


max119891119895notin119878119898minus1

[[Φ (119891119894 119888) minus 1


Φ(119891119894 119891119895)]] (13)


4 Classification













Complexity 9














7439











10 Complexity

6951

7698

6693

7302

SVM KNN CT ANN60006200640066006800700072007400760078008000

()








Acknowledgments


References















Complexity 11


































12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








Complexity 7


119867(119883) = sum119909isin120594

119901 (119909) log119901 (119909) (9)



CD = lim119903rarr0

( ln119862 (119903)ln 119903 ) (10)







Linear features





max119863 (119878 119888) = 1|119878| sum119891119894isin119878

Φ(119891119894 119888) (11)



min119877 (119878) = 1|119878|2 sum119891119894 119891119895isin119878

Φ(119891119894 119891119895) (12)


8 Complexity


max119891119895notin119878119898minus1

[[Φ (119891119894 119888) minus 1


Φ(119891119894 119891119895)]] (13)


4 Classification













Complexity 9














7439











10 Complexity

6951

7698

6693

7302

SVM KNN CT ANN60006200640066006800700072007400760078008000

()








Acknowledgments


References















Complexity 11


































12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








8 Complexity


max119891119895notin119878119898minus1

[[Φ (119891119894 119888) minus 1


Φ(119891119894 119891119895)]] (13)


4 Classification













Complexity 9














7439











10 Complexity

6951

7698

6693

7302

SVM KNN CT ANN60006200640066006800700072007400760078008000

()








Acknowledgments


References















Complexity 11


































12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








Complexity 9














7439











10 Complexity

6951

7698

6693

7302

SVM KNN CT ANN60006200640066006800700072007400760078008000

()








Acknowledgments


References















Complexity 11


































12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








10 Complexity

6951

7698

6693

7302

SVM KNN CT ANN60006200640066006800700072007400760078008000

()








Acknowledgments


References















Complexity 11


































12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








Complexity 11


































12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








12 Complexity





































Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018








Complexity 13































Journal of












Journal of







Volume 2018






Volume 2018















Journal of












Journal of







Volume 2018






Volume 2018








a pervasive approach to eeg-based depression detection€¦ · complexity psychiatrist....

Documents