eeg signals analysis using multiscale entropy for depth of

Research ArticleEEG Signals Analysis Using Multiscale Entropy forDepth of Anesthesia Monitoring during Surgery throughArtificial Neural Networks

Quan Liu1 Yi-Feng Chen2 Shou-Zen Fan3 Maysam F Abbod4 and Jiann-Shing Shieh5

1Key Laboratory of Fiber Optic Sensing Technology and Information Processing Ministry of EducationWuhan University of Technology Wuhan Hubei 430070 China2School of Information Engineering Wuhan University of Technology Wuhan Hubei 430070 China3Department of Anesthesiology College of Medicine National Taiwan University Taipei 100 Taiwan4College of Engineering Design and Physical Sciences Brunel University London Uxbridge UB8 3PH UK5Department of Mechanical Engineering and Innovation Center for Big Data and Digital Convergence Yuan Ze UniversityTaoyuan Chung-Li 32003 Taiwan

Correspondence should be addressed to Jiann-Shing Shieh jsshiehsaturnyzuedutw

Received 15 June 2015 Revised 23 August 2015 Accepted 7 September 2015

Academic Editor Dong Song

Copyright copy 2015 Quan Liu et alThis is an open access article distributed under theCreative CommonsAttribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

In order to build a reliable index to monitor the depth of anesthesia (DOA) many algorithms have been proposed in recentyears one of which is sample entropy (SampEn) a commonly used and important tool to measure the regularity of data seriesHowever SampEn only estimates the complexity of signals on one time scale In this study a new approach is introduced usingmultiscale entropy (MSE) considering the structure information over different time scales The entropy values over different timescales calculated throughMSE are applied as the input data to train an artificial neural network (ANN)model using bispectral index(BIS) or expert assessment of conscious level (EACL) as the target To test the performance of the new indexrsquos sensitivity to artifactswe compared the results before and after filtration by multivariate empirical mode decomposition (MEMD)The new approach viaANN is utilized in real EEG signals collected from 26 patients before and after filtering by MEMD respectively the results showthat is a higher correlation between index from the proposed approach and the gold standard compared with SampEn Moreoverthe proposed approach is more structurally robust to noise and artifacts which indicates that it can be used formonitoring the DOAmore accurately

1 Introduction

Anesthesia is an indispensable stage for doctors duringsurgery and in the intensive care environment which enablesthe patients to undergo surgery to keep unconsciousnessand lack of pain through suppressing response of nervoussystem to nonnoxious stimuli [1ndash3] However interaction ofanesthetic drugs and central nervous system is very complexso methodologies for assessment of DOA are controversialbut very important in medical domain [4ndash6] Monitoringthe DOA is not only to determine the patientsrsquo states duringsurgery but also to further control the amount of anestheticrequired for individuals to ensure high quality and safety

of anesthesia with rapid recovery after operation Thereforethe necessity to evaluate and optimize DOA monitoring isabsolutely important not only for surgeons during surgery butalso for patientsrsquo health after operation

In traditional methods measurement of DOA is imple-mented by analysis of signals collected from patients such aselectrocardiogram (ECG) respiration (Resp) blood pressure(BP) and peripheral oxygen saturation (SpO

2) which reflect

the consciousness level of patients indirectly However thesesignals cannot estimate the DOA accurately and are easilydisturbed by artifacts and noise EEG signal and auditoryevoked potential (AEP) based monitors are the internation-ally recognized anesthesia monitoring method in operation

Hindawi Publishing CorporationComputational and Mathematical Methods in MedicineVolume 2015 Article ID 232381 16 pageshttpdxdoiorg1011552015232381

2 Computational and Mathematical Methods in Medicine

[7 8] In particular the methods based on EEG for DOAevaluation have been developed rapidly The EEG signalswhich reflect the brainrsquos activities have been widely usedfor research and diagnosis especially for measuring theawareness level of patients EEG referring to brainrsquos electricalactivity is commonly recorded in a noninvasive approachwhich provides an available tool to study the human brainfor researchers and doctors [9] It has been widely usedfor measuring consciousness level of patients in medicalenvironment [10ndash12]

There are variousmethods based on EEG analysis appliedto monitor DOA recentlyThe bispectral index (BIS) monitorintroduced by Aspect Medical Systems Inc in 1994 [13ndash15] is widely used in the operation room for evaluatingthe DOA by analysis of EEG signals of patients duringsurgery BIS monitor has been proved as a reliable systemto measure the DOA except for several anaesthetic agentsin many researches [16 17] However the company thatintroduced the BIS monitor has not disclosed the detailedalgorithms In addition entropy monitors developed byDatex-Ohmeda produce response entropy (RE) and stateentropy (SE) to evaluate the irregularity in EEG signalsfor determining the DOA [18] The algorithm applied inthe Datex-Ohmeda entropy module calculates the RE andSE based on frequency domain approach called spectralentropy which is obtained by applying Shannon entropyto the power spectrum [19] However application of fastFourier transform (FFT) to estimate power spectrum maymiss the nonlinear and nonstationary properties of EEGsignals Although these two monitor systems are the mostpopular there are limitations Therefore an open sourceand time domain based method taking the nonlinear andnonstationary properties of EEG signals into considerationis need for monitoring DOA during surgery robustly andaccurately

The approximate entropy (ApEn) [20] and SampEn [21]algorithms are two powerful approaches proposed in appli-cation of determining the complexity of any time seriesAnd SampEn has been proved to perform better than ApEnfor monitoring DOA of patients during surgery in previousstudies [21ndash23] Nevertheless SampEn measures complexityof time series based on a single time scale so that it missesthe features associated with signal structure To overcomethis problem Costa et al introduced an improved methodnamedmultiscale entropy (MSE) to analyze the complexity ofbiological signals over multiple time scales [24 25] The EEGreflects the summation of human brainrsquos activity and containsthe information about neuronal dynamics underlying highand low frequency [26 27] Therefore MSE is appropriatefor obtaining the dynamics features related to multiple timescales and has been widely used in analysis of EEG recordings[28ndash30] Although MSE measure can explore the degree ofcomplexity over different time scales a single index whichindicates the DOA of patients during surgery is needed bysurgeons However we find that on the one hand manystudies applied MSE to distinguish the complexity of EEGthrough plotting entropy values over different time scalesoverall without considering time [31 32] or calculate entropyvalues on all time scales independently for monitoring DOA

[30] which is too complicated for surgeons to determinepatientsrsquo anesthesia level On the other hand in previousresearch [33] a single index was derived from entropy valuesbased on appropriate scales by averaging the scale dependententropies The limitation of this method is that entropiesrelated to each scale contribute unequally to measure thecomplexity and it is difficult to confirm the weights for eachindependent scale Therefore in this paper a new index isobtained from MSE analysis by combining the independententropies via ANN for measuring the DOA during surgery

TheANN is an extremely important and useful algorithmin machine learning inspired by biological neural networks[34] It can be used to adaptively and optimally estimatethe weights and functions which are generally unknown inadvance to depend on the input and target data by trainingvalidating and testing Therefore it has been widely used tosolve many tasks for classification and regression analysis inbiomedical engineering [35]

In this study simulated EEG corrupted with EOG andEEG collected from patients with different consciousnesslevel are analyzed by MSE to investigate the sensitivity toEOG and ability to distinguish the patientsrsquo states of entropiescorresponding to each independent scale Next we apply theMSE method to real EEG recordings collected from patientsduring surgery And then entropies and a gold standardare defined as input and target variables to train the ANNmodelThe outputs of ANN are severed as the new combinedindex for DOA monitoring BIS as a commercial index hasbeen approved by US Food and Drug Administration andmost widely used in operation room during surgery and ICUto monitor DOA although it is not perfect For exampleintraoperative awareness can occur during general anesthesiawith a small probability event even if BIS value is under60 according to recent researches [15 36 37] BIS is one ofthe technologies to accurately monitor the hypnotic effectsof general anesthetics and sedatives based on EEG signalsHowever the device is very expensive and the details of thealgorithms to calculate BIS index have not been disclosedSo it is necessary to create an open sources method forDOA monitoring accurately In this paper our aim is tocreate a new index which can accurately trace the changeof consciousness level of patients like BIS therefore BIS isused as a gold standard of DOA However it would be moreapplicable and reasonable if there is a real gold standard ofDOA as the target So in comparison ldquothe state of anestheticdepthrdquo called expert assessment of conscious level (EACL)[38] which is decided by five experienced anesthesiologistsbased on detailed recordings during surgery was used as thetarget to train ANN

EEG signals are always corrupted by artifacts such asEOG and EMG Generally the amplitude of EEG can beextremely less than artifacts so techniques are needed toremove EEG contaminants for accurate analysis In thispaper MEMD [39] based filter was used to remove artifactsfrom contaminated EEG signals On the one hand throughcomparison of performance of proposed method before andafter filtering we can indicate the robustness of proposedmethod to artifacts On the other hand MSE measurescomplexity of time series at different scales and filter can

Computational and Mathematical Methods in Medicine 3

2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(a)

2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(b)

2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(c)

2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(d)

Figure 1 EEG signal under different conditions (a) Original EEG collected from one patient under surgery (b) EEG after filtering of (a) bysumming IMF2 and IMF3 through MEMD (c) EEG corrupted with 5 dB EOG of (b) (d) EEG corrupted with minus5 dB EOG of (b)

enhance the features related to some scales for monitoringDOA more accurately If we combine the MSE at these scalethe indicator would have higher performance to measureDOA It is indicated that the index is less sensitive to noiseand perform highly better than SampEn

2 Materials and Methods

21 Data Sources and EEG Recordings The one channel EEGsignals tested in this study are collected from twenty-sixpatients through a foreheadmounted sensor byMP60 system(Philip IntelliVue MP60 BIS module) They aged from 23 to72 years are accepting ear nose and throat (ENT) surgerywith general anesthesia at the National Taiwan UniversityHospital (NTUH) of Taiwan when recording EEG And thedrugs administered for anesthesia induction and correspond-ing anesthetic technique are sevoflurane or desflurane fortracheal intubation of 18 patients sevoflurane or desfluranefor laryngeal mask airway (LMA) of 5 patients and propofolfor total intravenous anesthesia of 3 patients respectivelyThesampling rate of EEG is 125Hz

22 Data Preprocess According to the standard operationprocedure (SOP) with general anesthesia it can be dividedinto four stages that is the preoperation induction main-tenance and recovery [30] The collected EEG are dividedinto three parts in this study due to different purposes Firstlywe select ten patients at random to estimate the sensitivity of

MSE from each independent scale to EOG noise During thepreoperation stage patients prepare to accept the operationwith consciousness and always blink their eyes frequentlyso the EEG recordings during this stage are badly corruptedwith EOG artifact as shown in Figure 1(a) The EEG signalsduring preoperation stage are collected from these selectedten EEG recordings and then filtered using MEMD methodas the clean EEG signals as shown in Figure 1(b) Next we addEOG noise to the clean data with different noise signal ratio(SNR) ranging from 10 dB to minus20 dB with a step of minus1 dB withrespect to the EEG level as the simulated EEG data corruptedwith EOG artifact Figures 1(c) and 1(d) show the simulatedEEG data corrupted with two different EOG levels Secondlyonce again ten patients have been selected at random to assessand compare the ability of entropies from each independentscale to distinguish the patientsrsquo states under consciousnessor anesthesia So EEG data during preoperation and main-tenance stages are collected and filtered as mentioned aboveFinally all 26 patients are used to obtain a new single indexreflecting the DOA fromMSE via ANNmethod

23 Expert Assessment of Conscious Level Firstly tworesearch nurses keep observing the state of patients andrecording the events and signs which happen during surgeryin operation room and possibly have relationship with ldquothestate of anesthetic depthrdquo in detail and carefully [38] forexample the start and end time of the anesthetic eventsincluding induction and extubation drugs administered time


and their dose MAC values recorded every five minutesduring the whole period of anesthesia and so on Thenfive experienced anesthesiologists need to make decision bythe individual to plot the changes of ldquothe state of anestheticdepthrdquo of patients over the whole duration of operationbased on anesthesia record and their previous experiences Acontinuous curve was provided by each doctor to representhow deep under anesthesia the patient is In order to beconsistent with BIS the range of these curves is from 0 to100 and 100 indicates totally awake state and 0 is equivalentto EEG silence and a value between 40 and 60 representsan appropriate anesthesia level during surgery for generalanesthesia Because the original curve was plotted by handdrawing so finally it is digitalized and resampled with afrequency of 02Hz like BIS index to a single dimensionlessnumber series called expert assessment of conscious level(EACL) [38] Each anesthesiologist with different experiencemay have a different perspective on EACL therefore in orderto measure consciousness level more accurately the meanvalues of EACL from five anesthesiologists were obtainedas target instead of BIS index Figure 2 gives an exampleof EACL from five doctors Because these five doctorshave worked as anesthesiologists specially trained to giveanesthesia for many years the EACL they plotted based onanesthesia recordings and their experiences could be used asa real gold standard of DOA [38]

24 Multivariate Empirical Mode Decomposition Based FilterMEMD proposed by Rehman and Mandic in 2010 [39] isan improved algorithm of empirical mode decomposition(EMD) which was introduced by Huang et al in 1998 [40] InEMD method a signal is decomposed with iterative processinto several ordered elements called intrinsic mode functions(IMF) ranging from high to low frequency [41] In compari-son with conventional methods such as Fourier and Waveletdecomposition the EMD is driven by data adaptively withoutfixed basis functions So it is highly suitable for analyzingnonlinear and nonstationary signals And the original signal119883(119905) can be reconstructed by summing up all IMFs as follows

119883 (119905) =

119873

sum

119894=1

119888

119894(119905) + 119903

119873(119905)

(1)

where 119873 is the total number of IMFs decomposed by EMD119888

119894(119905) is the 119894th IMF and 119903

119873(119905) is the residue

The unwanted artifacts or noise can be removed byrecomposing the signal with different IMFs according to thefollowing equation

119883 (119905) =

119902

sum

119901

119888

119894(119905) (2)

where

119883(119905) is the filtered signal 119888119894(119905) is the 119894th IMF as

mentioned above and 119901 119902 isin (1119873) When 1 lt 119901 le 119902 =

119873 the signal is reconstructed with low frequency elementswhich means a low pass filter when 1 = 119901 le 119902 lt 119873 lowfrequency noise is removed which means a high pass filter

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(a)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(b)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(c)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(d)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(e)

Mean plusmn StdMean

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(f)

Figure 2 An example of EACL from five doctors (a) Doctor A (b)Doctor B (c) Doctor C (d) Doctor D (e) Doctor E (f) Mean andstandard error of EACL from five doctors

when 1 lt 119901 le 119902 lt 119873 it means a band pass filter and when119901 gt 119902 (2) can be expressed as follows

119883 (119905) =

119902

sum

1

119888

119894(119905) +

119873

sum

119901

119888

119894(119905) (3)

in this case it means a band stop filter According to analysisabove we can remove EOG from EEG signal by combiningthe selected IMFs which is signal dominated


MEMD

EEG

White noise

White noise

Filtered EEGIMF2 + IMF3

Figure 3 Schematic illustration of MEMD based filter

Although the MEMD is introduced to decompose multi-channel signals it also can be used in a single channel signalby combination of original signal and independent whitenoise added into extra channels to form amultivariate signalBy this means MEMD unlike ensemble empirical modedecomposition (EEMD) which decomposes white noiseadded signal and then averages out the noise by sufficientnumber of trials [42] solves the problem of mode mixingto a certain extent caused by EMD without introducingany white noise into original data [39] In comparison withEEMD MEMD introduces no noise and consumes less timewhen decomposing the signals In this paper MEMD isapplied to remove unwanted signals from EEG According tothe previous study [43] we reconstruct the EEG signals bysumming IMF2 and IMF3 after decomposition as the filteredsignals as shown in Figure 3

The conclusion that filtered EEG signals are reconstructedusing IMF2+ IMF3 is a statistically based empirically derivedby comparison of all possible combinations of IMFs fordiscriminating preoperation induction maintenance andrecovery stages and tracing the changes of consciousness levelin our previous study [43]Thirty patientsrsquo datawere collectedfor statistical analysis Firstly according to frequency rangesof the EEG signals IMF2 IMF3 IMF4 IMF5 and IMF6were considered for next combination So there totally are31 different ways Then due to the entropy values duringanesthesia state are less than awake state IMF2 IMF2+ IMF3IMF2 + IMF4 IMF2 + IMF3 + IMF4 and IMF2 + IMF3 +IMF6 were selected for the next analysis Finally 119901 valuesof entropy values were calculated to compare the statisticdifference between awake and anesthesia state and IMF2 +IMF3 with least 119901 value which is also less than 005 was usedas acceptable filtered EEG

25 Sample Entropy and Multiscale Entropy The SampEnis proposed by Richman and Moorman in 2000 [21] tomeasure the complexity of physical time series according tothe following steps

For a given time series with119873 points 119883(119894) 1 le 119894 le 119873embed dimension119898 tolerance 119903

(1) Form 119873 minus 119898 + 1 vectors 119883119898(119894) according to the

template defined as

119883

119898(119894) = 119909

119894+1 119909

119894+2 119909

119894+3 119909

119894+119898minus1

1 ⩽ 119894 ⩽ 119873 minus 119898 + 1

(4)

(2) Calculate the distance between two different vectorsmentioned above as

119889 [119883

119898(119894) 119883

119898(119895)] = max (10038161003816

1003816

1003816

119883

119898(119894 + 119896) minus 119883

119898(119895 + 119896)

1003816

1003816

1003816

1003816

)

for 0 ⩽ 119896 ⩽ 119898 minus 1(5)

(3) Count the total number of vectors 119883119898(119895) within 119903 of

119883

119898(119894) denoted by 119861

119894and then

119861

119898

119894(119903) =

1

119873 minus 119898 minus 1

119861

119894

119861

119898(119903) =

1

119873 minus 119898

119873minus119898

sum

119894=1

119861

119898

119894(119903)

(6)

(4) Set119898 = 119898 + 1 and repeat steps (1) to (3)

119860

119898(119903) =

1

119873 minus 119898

119873minus119898

sum

119894=1

119860

119898

119894(119903)

(7)

(5) Denote the SampEn by

SampEn (119898 119903119873) = minus ln 119860119898(119903)

119861

119898(119903)

(8)

Although SampEn is popular and useful in applicationof measuring complexity of signal it does not considerthe structure information related to time scales ThereforeCosta et al proposed MSE algorithm to analyze signals overdifferent scales [25] Firstly for a given scale 120591 a ldquocoarse-grainingrdquo process is made by averaging all the data pointslocated in a window which moves with step 120591 after which weget a new time series

119910

(120591)

119895=

1

120591

119895120591

sum

119894=(119895minus1)120591+1

119883

119894 (9)

Then SampEn algorithm is used for each new time series afterldquocoarse-grainingrdquo process related to time scale 120591 Figure 4shows the flow chart of coarse-graining procedure AndFigure 5 gives an example of MSE calculated from 30 simu-lated Gaussian white noise

In this paper the parameters are set as follows 120591 =

1 2 3 4 20119898 = 2 and 119903 = 02 according to the statisticalanalysis of previous studies [21 25 30]

26 Artificial Neural Network In this research a newmethodis proposed to obtain an index for monitoring DOA asshown in Figure 6 Figure 6(a) shows the detailed structureof ANN network and Figure 6(b) illustrates that structurefor each neuron 119882 = [1199081 1199082 1199083 119908119899] is the weightsof each neuron and 119887 is its bias In order to consider thestructure information related to multiple scales we measurethe complexity of EEG by MSE analysis Then the multiplescale entropies are transformed into a single index usingnonlinear regression method (eg ANN) to build the func-tions between MSE and the gold standard Generally an


120591 = 1

120591 = 2

120591 = 3

120591 = i

x2x1 x3 x4 x5 x6 xi xi+1 xi+2



y1 =x1 + x2

2y2 =

x3 + x42

y3 =x5 + x6

2yj =

xi + xi+12

y1 =x1 + x2 + x3

3y2 =

x4 + x5 + x63

yk =xi + xi+1 + xi+2

3

y1 =x1 + x2 + x3 + x4 + x5 + x6 + middot middot middot + xi

i

middot middot middot











Figure 4 Illustration of the coarse-graining procedure for MSE calculation Adapted from [25]

05

1

15

2

25

Sam

pEn

5 10 15 200Scale factor

Figure 5 MSE analysis of 30 simulated Gaussian white noise Eachnoise contains 30000 data points Symbols represent mean values ofentropy for the 30 time series and error bars the SD Adapted from[24 25]

integrated ANN model contains input layer hidden layerand output layer In this paper the input layer consists of 119873neurons ranged from 1 to 20 consistent with the number ofinputs hidden layer contains 20 neurons and output layerhas 1 neuron respectivelyThe target data is one-dimensionalseries regardless of the number of inputs We choose feed-forward backpropagation which is a very common methodto train ANN model as the learning rule Then the entropiesof all time scales calculated from EEG and gold standard aretreated as the input data and target data to train validate andtest theANNmodelThere aremultiple inputs to this networkand 1 output In order to confirm the performance of the newcombined index we also compare it with the entropy resultsrelated to a single scale from scale 1 to scale 20 via ANNIn this situation the input data of ANN is entropy valuesfrom a single scale Furthermore the samples percentages

divided randomly for training validation and testing are70 15 and 15 respectively All analyses were performedin MATLAB (v713 MathWorks Inc USA)

3 Results

In this section we compared the sensitivity of all entropyindexes of each independent time scale fromMSE analysis ofsimulated EEG corrupted with different level EOG artifactAnd then we analyzed the ability of each single scale entropyto distinguish the consciousness and anesthesia states ofpatient during surgery Finally the proposed method isapplied to real EEG signals collected from patients

31 Sensitivity of Single Scale Entropy to EOG In this sectionthe signals are used to evaluate the sensitivity of single scaleentropy to EOG artifacts The target of the filtering is toremove artifacts in EEG signals Through adding EOG noisewe simulated the contaminated EEG with different noiselevel then coefficient variation (ie the ratio of the standarddeviation to the mean CV) of MSE at each time scale isstatistically analyzed to compare their robustness to noise

The EEG signals collected from ten patients under pre-operation are used after filtration by summing IMF2 andIMF3 based on MEMD algorithm [43] Then the EOG as thesimulated artifact is added into the filtered EEG of each casewith different SNR ranging from 10 dB to minus20 dB with thestep of minus1 dB Considering the original filtered EEG there is32 different levelsrsquo signal plus the original filtered EEG foreach case A sliding window with 30-second length including3750 data points is utilized when measuring the complexityof EEG signals using MSE analysis and moves forward onceevery five seconds for real time DOA monitoring The CV ofthe entropy index for each single scale to the EOG artifactare analyzed We also plot the mean and standard deviationfor ten cases as indicated in Figure 7 We can see that the CVdecreases with the increasing of scales until scale 14 and then


Input layer Hidden layer

N

ANN

Network outputs

Input data (MSE)

Target data

Scale 1

Scale 2

Scale 3

Scale N

Output layer

20 1

middot middot middotmiddot middot middotmiddot middot middot

050100

0

1

2

0

1

2

0

1

2

0

1

2

050100

(a)

+ f

w1

w2

w3

wn

Input 1

Input 2

Input 3

Input n b

Output

(b)

Figure 6Theflowchart and structure of ANN (a)Theflowchart of proposedMSE basedmethod viaANNand the structure of ANNnetwork(b) The structure of single neuron in ANN for each layer

5 10 15 200Time scales

0

01

02

03

04

CV

Figure 7Themean and standard deviation of CV of entropy valuesfor 32 different levelsrsquo EOG noise Symbols represent mean values ofCV from ten cases with different level of noise and error bars the SD

it rises slightly but extremely less than the value of scale 1Theresults indicate that the entropy at scale 1 is the most sensitiveto EOG artifactThe possible reason is that entropy values onsmall time scales mainly represent the information of highfrequency parts And values on large time scales indicatethe low frequency information due to the ldquocoarse-grainingrdquoprocess which averages the data points within a fixed-sizewindow so that the high frequency parts are removed from

original signals [24 32] With the ldquocoarse-grainingrdquo processthe amplitudes of EOG decrease so EOG have less effect onEEG signals It is noted that MSE at scale 1 are the mostsensitive to EOG artifact which means that SampEn is proneto artifacts So combination of MSE at multiple time scaleinstead of SampEn may provide a more robust method tomonitor DOA

32 Comparative Study of Each Scale for DistinguishingDifferent States In order to further evaluate the ability ofMSE at different time scales to distinguish the patientsrsquo statesduring surgery we test the MSE on real EEG signals col-lected from ten patients under preoperation andmaintenancestages before filtering Moreover we investigate the effect offiltering by summing IMF2 and IMF3 [43] on MSE at eachindependent scale As shown in Figure 8 the error bar ateach scale represents the mean and standard deviation of anentropy measured from ten patients For time scale one themean value of entropy is very close between preoperation andmaintenance before filtering which indicates that it is difficultto differentiate these two stages using MSE at scale factor ofone With the increasing of time scales the entropy valuesfrom stage 1 decrease significantly and then rise slightly butvalues from stage 3 ascend extremely and then decline slowly


Stage 1Stage 3

5 10 15 200Scale factors

0

05

1

15

2

25En

tropy

val

ues

(a)

Stage 1Stage 3


0

05

1

15

2

25

Entro

py v

alue

s(b)

Figure 8 Statistical analysis of MSE from ten patientsrsquo EEG under different states at multiple time scales Stage 1 is preoperation and stage 3is maintenance during surgery Symbols marked with ldquolowastrdquo indicate significant difference between stage 1 and stage 3 with 119901 lt 005 (a) Beforefiltering (b) After filtering

Then the paired-samples two-tailed 119905-test was used tocompare the difference of MSE from ten patientsrsquo EEGbetween stage 1 and stage 3 at multiple time scales before andafter filtering The level of significance was set at 119901 lt 005

as shown in Figure 8 At all scales except scale 1 the meansof the entropy values have significant difference betweenthese two stages before filtering The second figure shows theresults from analysis after filtering the means of the entropyvalues at all scales are statistically significantly different Thefilter performs usefully at scale 1 Although the difference ofentropy between stage 1 and stage 3 decreases at large timescales regardless of being filtered or not MSE at large scalesalso have the capability of differentiating stage 1 and stage 3with 119901 lt 005 We can see that the entropy value from stage 1is largely outweighing the value from stage 3 at the scale of onewhich means that EEG from patients under consciousnessperform with more complexity than under anesthesia stateTherefore entropy at scale 1 can distinguish the EEG collectedfrom patients under these two stages after filtering with 119901 lt005 It is consistentwith SampEnwhich is equal to theMSE ata time scale of one tomonitor DOAduring surgery accordingto previous researches [22 30 43] And then the entropyvalues from stage 3 surpass stage 1 which means that EEGfrom patients under maintenance state are more complexityin comparison with preoperation at large time scale

The entropy values reduce greatly for both stages inFigure 8(b) compared to Figure 8(a) when scale factorexceeds 3 It is possible because low frequency parts areremoved from original EEG signals by summing IMF2 andIMF3 so entropy values at large scales related to low fre-quency elements decrease a lot And filtering also changes theSD of time series which will affect the tolerance level Because

main element reserved SD remain largely unchanged com-pared to prefiltering in spite of slight decrease while withthe loss of high frequency elements due to coarse-grainingprocedure the amplitude decreases Therefore slight changein the tolerance level compared with relative lower amplitudeindicates that fewer vectors will be distinguishable and thatcomplexity of signal will decrease [24] When comparingthese two figures the entropy value from stage 1 after filteringis larger than before filtering at time scale 1 because EOGartifacts which are of low frequency and relatively lesscomplexity are canceled from EEG The entropy at scale 1 isextremely sensitive to EOG artifact

By the above analysis entropy at all scales can make acontribution to discriminate the EEG during preoperationand maintenance stages although the ability of MSE at somescales is very weak no matter before or after filtering And itis not robust to monitor DOA using SampEn which is equalto MSE at scale 1

33 Performance Evaluation for Monitoring DOA In thissection 26 patientsrsquo EEG signals collected during surgery areused to investigate the performance of our proposed methodto monitor DOA A sliding window with 30-second lengthincluding 3750 data points is utilized when measuring thecomplexity of EEG signals using MSE analysis and movesforward once every five seconds for real time DOA moni-toring The prediction of ANN is quantified by coefficient ofdetermination denoted by 1198772 a measure of the proportionof total variation of network outputs is replicated by ANNmodel 1198772 is larger when prediction value of ANN is closerto target data Furthermore correlation coefficient (Cor-rcoef) is employed to measure the linear correlation between


Before filteringAfter filtering

0

02

04

06

08

1C

orre

latio

n co

effici

ents

(Cor

rcoe

f)


(a)


0

02

04

06

08

1

Cor

relat

ion

coeffi

cien

ts (C

orrc

oef)


(b)

Figure 9 Correlation coefficients of MSE and gold standard at different time scale before and after filtering Symbols represent mean valuesof correlation coefficients between MSE based measurement and gold standard from 26 cases and error bars the SD (a) BIS index was usedas gold standard (b) EACL was used as gold standard

the new index obtained from MSE via ANN and goldstandard to confirm the accuracy and robustness of proposedmethod for DOA monitoring

As shown in Figure 9 there presents a moderate linearrelationship between MSE at each scale from 1 to 20 andgold standard before filtering and a moderate or even weaklinear relationship after filtering The correlation coefficientsproduced at large scales appear to be consistent with theanalysis mentioned above which indicates that the MSE atlarge scales are less capable of tracking the consciousness levelof the patientsThe119877 values of ANNmodel are relatively lowsince the relationship between MSE at single scales and goldstandard is not so strong the ANN model misses entropypoints by much Furthermore the correlation coefficientat scale 1 after filtering is higher than before filtering Itdemonstrates that the filtering algorithm used in this studyis most effective for MSE at scale 1 (ie SampEn) as indicatedin Figure 9 consisting of the error bars which represent themean and standard deviation but remove lots of informationrelated to large scales Therefore the mean and standarddeviation of correlation coefficient and 119877 value after filteringare smaller in comparison with those before filtering Andthe large values of standard deviation in Figure 9 suggestthat MSE at single scales is extremely sensitive to noise formonitoring DOA

Tables 1 and 2 show results focused on combinations bychanging scales from 1 to 20 from MSE of EEG to forma composite indicator for measuring DOA before and afterfiltering For example 1-1 representing the input data of ANNmodel is MSE at only time scale 1 and 1ndash20 indicates thatthere are 20 scales from 1 to 20 used to train the ANN

model as inputs and so on We can note that both 119877 valuesand correlation coefficient increase generally in spite of somefluctuationwith addingmore entropies at different scales intothe network as inputs Indeed by the above analysis theMSEat different scales make contribution to track the anesthesialevel from EEG analysis So combination of multiple scalescan enhance this feature and perform better to measure theDOA than single scale Furthermore it is indicated thatCV decrease with adding more scales The lower CV valuesuggests that the corresponding index performs less sensitiveto noise because the means of the entropy values at scalefactor 2ndash20 are statistically significantly different no matterbefore or after filtering as shown in Figure 8 So it is uncertainthat the nonfiltering produces better results than filtering ifscale factors 2ndash14 are selected But it is confirmed that ifall scales ranged from 1 to 20 are used there will be betterresults MSE at scale from 2 to 20 before and after filteringhave similar performance but worse than MSE at scale from1 to 20 as shown in Tables 1 and 2 Furthermore the similarperformance between before and after filtering indicates thatMSE based index via ANN is robust to noise

Based on the results above entropies at all scales from 1to 20 are used to train ANN model as input data to obtaina composite indicator for DOA monitoring Tables 3 and 4show the results of 26 patients using our proposed method tomonitor DOA using BIS and EACL as target during surgeryMSE based index via ANN appears to be a very strongpositive correlationwith the gold standard and thus performsextremely better compared with SampEn for monitoringDOA during surgery Moreover it is evident in Figure 9 thatthe correlation between MSE at scale factors of 1 4 and 5


Table 1 Different combinations of multiple scales changing from 1 to 20 via ANN for monitoring DOA using BIS as gold standard 1-1 in thefirst column represents the input data of ANNmodel that is the entropy at time scale 1 and 1ndash20 indicates that there are multiple scales from1 to 20 and so on

Scale(s) Before filtering After filtering119877 Corrcoef CV () 119877 Corrcoef CV ()

1-1 044 040 plusmn 021 5250 053 051 plusmn 023 45101-2 047 050 plusmn 027 5400 068 070 plusmn 019 27141ndash3 067 068 plusmn 019 2794 069 070 plusmn 018 25711ndash4 069 070 plusmn 014 2000 072 072 plusmn 015 20831ndash5 069 070 plusmn 014 2000 072 072 plusmn 015 20831ndash6 071 071 plusmn 014 1972 073 073 plusmn 015 20551ndash7 071 071 plusmn 014 1972 073 073 plusmn 014 19181ndash8 071 071 plusmn 014 1972 074 073 plusmn 014 19181ndash9 072 072 plusmn 014 1944 075 074 plusmn 014 18921ndash10 071 071 plusmn 016 2254 075 074 plusmn 014 18921ndash11 072 072 plusmn 014 1944 076 075 plusmn 013 17331ndash12 075 074 plusmn 012 1622 076 075 plusmn 014 18671ndash13 073 073 plusmn 014 1918 076 075 plusmn 013 17331ndash14 076 074 plusmn 012 1622 076 075 plusmn 013 17331ndash15 075 074 plusmn 013 1757 076 074 plusmn 014 18921ndash16 075 074 plusmn 013 1757 076 075 plusmn 013 17331ndash17 076 074 plusmn 013 1757 077 075 plusmn 013 17331ndash18 076 074 plusmn 013 1757 075 075 plusmn 014 18671ndash19 076 074 plusmn 013 1757 076 075 plusmn 012 16001ndash20 076 075 plusmn 012 1600 076 075 plusmn 015 20002ndash20 074 073 plusmn 013 1781 073 072 plusmn 013 1806Mean plusmn std 070 plusmn 009 070 plusmn 009 2230 plusmn 1061 073 plusmn 005 073 plusmn 005 2063 plusmn 621

and BIS is higher after filtering than that before filteringthus a feasible optimization method is built by selectingentropies at which scales perform better before and afterfiltering to retrain the ANN model and acquire a new indexlisted in the final column in Table 3 That is to say entropiesat the scale factors of 1 4 and 5 after filtering and otherfactors before filtering are used to train the ANN as inputsThe correlation is higher between MSE from combinationof pre- and postfiltering and BIS compared with SampEnand MSE from pre- and postfiltering respectively BesidesCV of MSE from combination of pre- and postfiltering (ie964) is extremely smaller than SampEn (ie 5250 and 4510)and MSE (ie 1600 and 2000) before and after filteringrespectively It indicates that optimization method performsmore accurate robustness to monitor the DOA as a newindicator

Tables 3 and 4 show the statistic results of MSE basedmeasurement via ANN with EACL and BIS as gold stan-dard The proposed MSE based method not only has highcorrelation with BIS index but also is very similar to EACLIt indicates that this method is successful in measuringthe consciousness level and monitoring DOA Furthermorethe proposed method using EACL as target performs moreaccurately with higher correlation compared with BIS It isknown that BIS is prone to artifacts while the proposedmethod based on MSE and ANN is extremely robust due tothe high similarity to gold standard (ie EACL and BIS) no

matter before or after filteringMoreover BIS index have beenquestioned for its reliability to monitor DOA so using EACLas gold standard would be more acceptable

Furthermore in this paper 10 patients received desflu-rane 13 patients received sevoflurane and 3 patients receivedpropofol as anesthesia agents Table 5 presents the meanand standard deviation of correlation coefficient betweenMSE via ANN and EACL for monitoring DOA in terms ofpropofol sevoflurane and desflurane respectively There is ahigh correlation for each agent especially MSE via ANN bycombination of pre- and postfiltering Therefore there is nodifference between propofol sevoflurane and desflurane forDOA monitoring

4 Discussion

SampEn is a method widely used in many researches tomeasure complexity of signals and monitor DOA duringsurgery [21 22 43] MSE an improved algorithm fromSampEn measures complexity of signal at different timescales and is also commonly applied to complex physiologicaltime series [24 25 29 30 32] However a single indexneeded from the MSE analysis for monitoring DOA andrelative complexity at multiple scales must be taken intoaccount in clinical applications ANN which can adaptivelyand optimally evaluate the function between MSE and asingle index depending on the input and target data by


Table 2 Different combinations of multiple scales changing from 1 to 20 via ANN for monitoring DOA using EACL as gold standard 1-1 inthe first column represents the input data of ANN model that is the entropy at time scale 1 and 1ndash20 indicates that there are multiple scalesfrom 1 to 20 and so on



training validating and testing provide a special solution tothis task When they were taken into account independentlyall the scales are analyzed to confirm the sensitivity to noiseand contribution to strengthen the indicatorrsquos preciseness forprediction DOA

EOG artifact which is the most common noise in EEGis added to original signals with different level The resultsshow that MSE at the scale factor of one is more sensitive tonoise with high CV which is also proved in 26 patientsrsquo realEEG signals MSE are calculated using SampEn algorithmafter coarse-graining procedure The accuracy of SampEndepends on time series length [21]Thediscrepancies betweenSampEn values numerically increasewith the decrease of datalength [24] while this coarse-graining procedure decreasesthe number of data points with the increase of time scaleAlthough it is uncertain of the minimum length of datarequired to calculate MSE the error because of decreaseddata number will increase [24] In this paper the windowsize for MSE calculation is 3750 There are less than 250 datapoints when scale factors are more than 15 The consistencyof SampEn is extremely decreases So CV increases again inthe last few scale factor as shown in Figure 7

And then we investigate the ability to distinguish andtrack the change of anesthetic states MSE at scale 1 per-form better after filtering than before filtering but filteringalgorithm removes lots of information associated with largetime scales in spite of filtering out physiological and external

noise effectively by summing IMF2 and IMF3 [43] We alsonote that filtering takes no effect on proposed method whenmeasuring anesthesia depth as shown in Tables 3 and 4 Thereason is that the ability of MSE at scale factor of one todistinguish patientsrsquo states increases after filtering in spite ofdecrease at larger scales in comparison with prefiltering asindicated in Figure 8 And standard deviations of MSE forstage 1 at each scale after filtering are smaller than thosebefore filtering Therefore the filtering takes little effect onoverall performance of proposed approach for monitoringDOA Finally in order to optimize the composite index theentropies at which scales perform better before and afterfiltering are selected to train ANN model The correlationbetweenMSE from combination of pre- and postfiltering andthe gold standard is highest comparedwith SampEn andMSEfrom pre- and postfiltering respectively The results confirmthat our proposed method is more accurate and robust tomeasuring DOA than SampEn

Generally the frequency of EEG signal can be dividedinto bands and the pattern within a certain frequency rangecontains the corresponding biomedical features So EEGfiltered with different passed band have different usefulcharacteristics for monitoring DOA Entropy monitoringcommercially developed by Datex-Ohmeda measures DOAof patient at two different frequency bands which producesresponse entropy and state entropy The combination ofdifferent parameters derived from multiple bands of EEG


Table 3 Combination of all scales ranging from 1 to 20 via ANN for monitoring DOA using BIS as gold standard SampEn as shown in 2thand 4th columns are the correlation coefficients between ANN outputs and BIS using MSE at the scale factor of one as training inputs beforeand after filtering The 3th and 5th columns are the results using MSE at scales from 1 to 20 as training inputs before and after filtering andthe final column is the corresponding results using MSE at scales 2 3 and 6ndash20 before filtering and 1 4 and 5 after filtering

Cases Before filtering After filtering Combination of pre- and postfilteringSampEn MSE SampEn MSE

1 074 087 092 090 0922 038 071 047 070 0863 065 090 051 090 0924 043 086 028 085 0925 085 087 084 091 0906 042 070 031 071 0747 049 073 051 070 0778 012 064 051 068 0799 048 083 074 082 09110 058 073 071 077 08111 000 079 034 085 08512 052 086 075 083 09013 044 085 068 085 09014 022 040 035 037 07215 058 087 067 084 09116 035 078 070 086 08517 028 068 068 075 07918 055 087 018 082 08719 019 079 067 082 08420 045 070 058 075 07521 031 068 049 078 08122 039 070 027 072 07623 032 051 032 033 06224 minus001 056 minus003 053 07025 021 075 059 072 08526 050 083 032 082 083Mean plusmn std 040 plusmn 021 075 plusmn 012 051 plusmn 023 075 plusmn 015 083 plusmn 008CV () 5250 1600 4510 2000 964119877 044 076 053 076 085

would provide more accurate knowledge for DOA moni-toring SampEn measures the complexity of EEG at singletime scale However MSE measures the complexity of timesseries at different time scales The features enhanced by filterused for optimizing SampEn may not be suitable for MSE atdifferent scales It is noted that there is a significant differencebetween stage 1 and stage 3 after filtering for MSE at scale1 but there is no difference before filtering as shown inFigure 8 So it is reasonable to consider MSE at scale 1 afterfiltering instead of before filtering as one of the inputs toobtain complex parameter via ANN For MSE at other scalesthough there is a significant difference between stage 1 andstage 3 no matter before or after filtering the correlationafter filtering between MSE at scale 3 for EACL and scales4 and 5 for BIS as target is higher than before filtering soMSE with better performance are chose to measure DOA Inthis way the complex index derived from MSE at multipletime scale with better performance via ANN would be more

accurate to characterize DOA In this paper the filtered EEGreconstructed by IMF2 + IMF3 are used to optimize MSEat scale 1 If reconstructed using different combination ofIMFs to optimize MSE at every other scale we combineall optimizing MSE at corresponding scale via ANN thenthe index calculated by proposed method would be moreaccurate forDOAmonitoring It is our nextwork in the futureto find the combination of IMFs to optimize MSE at everyother scale

In conventional methods time and frequency domainsanalyses of EEG signals are used to measure the conscious-ness level such as median frequency spectral edge frequencyspectral entropy and ApEn ApEn is a valuable method tocalculated complexity from a dynamical system in phasespace Because ApEn sets a threshold for noise cancellationit is better than conventional method in measuring the con-sciousness level from EEG recordings [20 44] FurthermoreSampEn is the improved algorithm fromApEn so it performs


Table 4 Combination of all scales ranging from 1 to 20 via ANN for monitoring DOA using EACL as gold standard SampEn as shown in2th and 4th columns are the correlation coefficients between ANN outputs and EACL using MSE at the scale factor of one as training inputsbefore and after filtering The 3th and 5th columns are the results using MSE at scales from 1 to 20 as training inputs before and after filteringand the final column is the corresponding results using MSE at scales 2 and 4ndash20 before filtering and 1 and 3 after filtering


1 060 085 080 089 0882 064 081 047 084 0853 073 088 024 086 0894 069 092 020 090 0935 083 090 088 091 0936 048 087 060 087 0897 054 090 066 090 0918 036 081 052 082 0869 042 085 056 081 09110 041 078 053 085 08011 019 078 023 085 08812 052 085 076 084 08913 061 090 078 090 09114 042 081 061 083 08615 057 087 070 086 09316 045 080 073 085 08917 032 071 067 079 08218 056 087 062 085 08919 023 081 070 084 08720 043 075 060 074 07921 042 073 051 081 08522 033 074 045 076 08223 042 078 070 079 08524 073 085 074 083 08825 048 084 067 087 08926 066 085 066 089 090Mean plusmn std 050 plusmn 016 083 plusmn 006 060 plusmn 017 084 plusmn 004 088 plusmn 004CV () 3140 689 2883 518 429119877 052 084 061 084 089

Table 5 The statistic results of performance on three different anesthesia agents

Anesthesia drugs MSE via ANN before filtering MSE via ANN after filtering Combination of pre- and postfilteringDesflurane 082 plusmn 005 084 plusmn 003 087 plusmn 002Sevoflurane 083 plusmn 005 085 plusmn 003 088 plusmn 005Propofol 086 plusmn 003 084 plusmn 004 090 plusmn 002

better than or as good as ApEn at least [21 22]The combinedindex behaves better than SampEn as reference performancemeasure confirmed by higher correlation with gold standardIn our opinion it is adequate to demonstrate the performanceof present index in this paper Nevertheless there are manynew methods applied to EEG signals to measure nonlinearand complexity Gifani et al proposed optimal fractal-scalinganalysis to quantify human EEG dynamic for depth ofanesthesia [45] In 2012 Kumar et al used fractal dimensionof EEG to assess hypnosis state of patient during anesthesia[46] And in 2015 Hayashi et al used poincare analysis in

estimating anesthesia depth [47] and Melia et al provided amethodology for prediction of nociceptive responses duringsedation to quantify analgesia level from EEG signals in highfrequencies [48]Theywere successfully applied inmeasuringthe depth of anesthesia based on EEG signal in specialaspect Futurework is needed to draw the conclusionwhetherour method performs better than these methods mentionedabove in detail It is difficult to say which method is best [49]on the one hand it is impossible to apply all these methodsto the current population under study And each methodalso can use different parameters and improved algorithms


For example a refined version of the MSE may furtherimprove the performance to a certain extent A detailedstudy of implication using different MSE methods will bedemonstrated in future On the other hand these methodsmay be tested under different condition and database Itis worth having an attempt to fuse multiple parametersextracted from EEG for a reliable monitor

In addition during deep anesthesia burst suppressionin EEG is recognised as light anesthesia which is a seriousproblem in EEG based indicators when other methods areused such as median frequency and spectral edge frequency[20 44] BIS can successfully avoid this problem by definitionof a burst suppression ratio [13] In present study there isno burst suppression component in collected EEG signalsthus this problem has no effect on our results FurthermoreApEn and SampEn can correctly monitor burst suppressionoccurring during deep anesthesia as anesthesia concentrationincreases according to previous studies [20 21 44] MSE asan advancedmethod improved from SampEn could calculatethe complexity of data series over different time scalesIt is reasonable to believe that EEG analysis using MSEfor DOA monitoring via regression with BIS can avoidmisclassification of burst suppression although more futurework is needed for confirmation

Propofol sevoflurane and desflurane are three com-monly used anesthesia agents for induction and mainte-nance of general anesthesia They have been known to havethe same mechanisms of action all through potentiationof GABAA receptor activity [50] EEG research finds thatthey cause a prominent decrease in gamma-band activityundergoing general anesthesia [51] Propofol may be thepreferred induction anesthetic for a shorter time surgerycomparedwith sevoflurane or desflurane [52]However thereare many previous studies to show that propofol does notshow a significant difference compared with sevofluraneor desflurane in patients undergoing surgery for anestheticinduction and maintenance [52 53] It is consistent with ourresults shown in Table 5 However more data is needed toconfirm this conclusion in the future

The purpose of this paper is to propose a new approachusing MSE via ANN and confirm its performance for mon-itoring DOA The parameters of ANN such as the neuronsnumber of each layer in ANN model seem to have less limi-tation on the results so the parameters are selected regardlessof optimization Nevertheless we will try to optimize all theparameters as far as possible in the next step for applicationto practice We note that MSE at large scales can makecontribution to track the change of consciousness level ofpatients in spite of being very weak and the correlationbetween composite index based on MSE via ANN and goldstandard strengthens with adding more scales to train themodel However the maximum scale ofMSE is set to be 20 inthis study Somore experiments are need to confirm the effectof scales larger than 20 Furthermore in order to integrateMSE over different scales into a single index we need toselect the appropriate scales In this paper we analyze parts ofvarious combinations and integrate MSE at all scales into thesingle indicator of anesthesia depthMore deliberate selectionof scale combinations are needed to be further explored

The EACL data are derived from five experienced anes-thesiologists through quantifying the consciousness levels ofeach patients according to operation recordings and theirexperience as the depth of anesthesia By this means thepresent method can avoid the problems occurring in BISand thus can be extended to other anesthesia techniquesHowever this mentioned method measures consciousnesslevel based on EEG signals generated by cerebral cortex likeBIS not all drugs administered for anesthesia act on this partFor example if they are acting on thalamus and brain stem[54] this method is not suitable in these cases

5 Conclusions

In this paper a new method is proposed to monitor DOAof patients during surgery based on MSE via ANN Itseffectiveness is evaluated by correlation analysis with BISThenew index performs extremely better than the raw single scaleMSE index and SampEn The index from MSE by combina-tion of pre- and postfiltering is the most accurate indicatorfor determining the DOA in patients during surgeryThere isa very strong positive correlation (ie 083 plusmn 008) betweenproposed index and BIS and a lower CV (ie 964) whichindicates that the new approach can be very useful foraccurate and robust measurement of DOA

Conflict of Interests

The authors declare that there is no conflict of interests

Acknowledgments

This research was financially supported by the Center forDynamical Biomarkers and TranslationalMedicine NationalCentral University Taiwan which is sponsored by Ministryof Science and Technology (Grant no MOST103-2911-I-008-001) Also it was supported byNational Chung-Shan Instituteof Science amp Technology in Taiwan (Grant nos CSIST-095-V301 and CSIST-095-V302) and National Natural ScienceFoundation of China (Grant no 51475342)

References

[1] I Rampil ldquoMechanisms and sites of action of general anaesthet-icsrdquo inMemory and Awareness in Anaesthesia IV Proceedings ofthe Fourth International Symposium on Memory and Awarenessin Anaesthesia p 223 World Scientific 2000

[2] J A Campagna K W Miller and S A Forman ldquoMechanismsof actions of inhaled anestheticsrdquo The New England Journal ofMedicine vol 348 no 21 pp 2110ndash2124 2003

[3] E R John and L S Prichep ldquoThe anesthetic cascade a theory ofhow anesthesia suppresses consciousnessrdquo Anesthesiology vol102 no 2 pp 447ndash471 2005

[4] T A Bowdle ldquoDepth of anesthesia monitoringrdquo AnesthesiologyClinics of North America vol 24 no 4 pp 793ndash822 2006

[5] G A Mashour ldquoIntegrating the science of consciousness andanesthesiardquo Anesthesia and Analgesia vol 103 no 4 pp 975ndash982 2006


[6] B A Orser ldquoLifting the fog around anesthesiardquo ScientificAmerican vol 296 no 6 pp 54ndash61 2007

[7] B Heyse B Van Ooteghem B Wyler M M R F Struys LHerregods and H Vereecke ldquoComparison of contemporaryEEG derived depth of anesthesia monitors with a 5 stepvalidation processrdquo Acta Anaesthesiologica Belgica vol 60 no1 pp 19ndash33 2009

[8] P S Myles ldquoPrevention of awareness during anaesthesiardquo BestPractice amp Research Clinical Anaesthesiology vol 21 no 3 pp345ndash355 2007

[9] E Niedermeyer and F L da Silva Electroencephalography BasicPrinciples Clinical Applications and Related Fields LippincottWilliams ampWilkins 2005

[10] L C Jameson andT B Sloan ldquoUsingEEG tomonitor anesthesiadrug effects during surgeryrdquo Journal of Clinical Monitoring andComputing vol 20 no 6 pp 445ndash472 2006

[11] X-S Zhang R J Roy and E W Jensen ldquoEEG complexity as ameasure of depth of anesthesia for patientsrdquo IEEE Transactionson Biomedical Engineering vol 48 no 12 pp 1424ndash1433 2001

[12] M Jospin P Caminal E W Jensen et al ldquoDetrended fluctua-tion analysis of EEG as a measure of depth of anesthesiardquo IEEETransactions on Biomedical Engineering vol 54 no 5 pp 840ndash846 2007

[13] J C Sigl and N G Chamoun ldquoAn introduction to bispectralanalysis for the electroencephalogramrdquo Journal of ClinicalMonitoring vol 10 no 6 pp 392ndash404 1994

[14] I Kissin ldquoDepth of anesthesia and bispectral index monitor-ingrdquo Anesthesia amp Analgesia vol 90 no 5 pp 1114ndash1117 2000

[15] M S Avidan L Zhang B A Burnside et al ldquoAnesthesiaawareness and the bispectral indexrdquo The New England Journalof Medicine vol 358 no 11 pp 1097ndash1108 2008

[16] Y Punjasawadwong A Phongchiewboon and N Bunchung-mongkol ldquoBispectral index for improving anaesthetic deliveryand postoperative recoveryrdquoThe Cochrane Database of System-atic Reviews no 4 Article ID CD003843 2007

[17] S D Whyte and P D Booker ldquoBispectral index during isoflu-rane anesthesia in pediatric patientsrdquo Anesthesia amp Analgesiavol 98 no 6 pp 1644ndash1649 2004

[18] G N Schmidt P Bischoff T Standl A Hellstern O Teuberand J Schulte ldquoComparative evaluation of the Datex-OhmedaS5 entropy module and the bispectral index monitor duringpropofol-remifentanil anesthesiardquo Anesthesiology vol 101 no6 pp 1283ndash1290 2004

[19] H Viertio-Oja V Sarkela M Sarkela et al ldquoDescription of theEntropy algorithm as applied in theDatex-Ohmeda S5 EntropymodulerdquoActa Anaesthesiologica Scandinavica vol 48 no 2 pp154ndash161 2004

[20] J Bruhn H Ropcke and A Hoeft ldquoApproximate entropy asan electroencephalographic measure of anesthetic drug effectduring desflurane anesthesiardquo Anesthesiology vol 92 no 3 pp715ndash726 2000

[21] J S Richman and J R Moorman ldquoPhysiological time-seriesanalysis using approximate and sample entropyrdquo AmericanJournal of PhysiologymdashHeart and Circulatory Physiology vol278 no 6 pp H2039ndashH2049 2000

[22] QWei Q Liu S-Z Fan et al ldquoAnalysis of EEG via multivariateempiricalmode decomposition for depth of anesthesia based onsample entropyrdquo Entropy vol 15 no 9 pp 3458ndash3470 2013

[23] X Chen I C Solomon and K H Chon ldquoComparison of theuse of approximate entropy and sample entropy applicationsto neural respiratory signalrdquo in Proceedings of the 27th Annual

International Conference of the Engineering in Medicine andBiology Society pp 4212ndash4215 IEEE September 2005

[24] MCosta A L Goldberger andC-K Peng ldquoMultiscale entropyanalysis of biological signalsrdquo Physical Review E vol 71 no 2Article ID 021906 2005

[25] MCosta A L Goldberger andC-K Peng ldquoMultiscale entropyanalysis of complex physiologic time seriesrdquo Physical ReviewLetters vol 89 no 6 Article ID 068102 2002

[26] G Buzsaki Rhythms of the Brain Oxford University PressOxford UK 2006

[27] R Scheeringa P Fries K-M Petersson et al ldquoNeuronaldynamics underlying high- and low-frequency EEGoscillationscontribute independently to the human BOLD signalrdquo Neuronvol 69 no 3 pp 572ndash583 2011

[28] T Takahashi R Y Cho T Mizuno et al ldquoAntipsychoticsreverse abnormal EEG complexity in drug-naive schizophreniaamultiscale entropy analysisrdquoNeuroImage vol 51 no 1 pp 173ndash182 2010

[29] G Ouyang C Dang and X Li ldquoMultiscale entropy analysisof EEG recordings in epileptic ratsrdquo Biomedical EngineeringApplications Basis and Communications vol 21 no 3 pp 169ndash176 2009

[30] Q Liu Q Wei S-Z Fan et al ldquoAdaptive computation of mul-tiscale entropy and its application in eeg signals for monitoringdepth of anesthesia during surgeryrdquo Entropy vol 14 no 6 pp978ndash992 2012

[31] T Mizuno T Takahashi R Y Cho et al ldquoAssessment of EEGdynamical complexity in Alzheimerrsquos disease using multiscaleentropyrdquoClinical Neurophysiology vol 121 no 9 pp 1438ndash14462010

[32] J-H Park S Kim C-H Kim A Cichocki and K Kim ldquoMul-tiscale entropy analysis of EEG from patients under differentpathological conditionsrdquo Fractals vol 15 no 4 pp 399ndash4042007

[33] D Li X Li Z Liang L J Voss and J W Sleigh ldquoMulti-scale permutation entropy analysis of EEG recordings duringsevoflurane anesthesiardquo Journal of Neural Engineering vol 7 no4 Article ID 046010 2010

[34] M T Hagan H B Demuth and M H Beale Neural NetworkDesign PWS Publisher Boston Mass USA 1996

[35] R Nayak L Jain and B Ting ldquoArtificial neural networks inbiomedical engineering a reviewrdquo in Proceedings of the Asia-Pacific Conference on Advance Computation 2001

[36] T JMcCulloch ldquoUse of BISmonitoring was not associated witha reduced incidence of awarenessrdquo Anesthesia amp Analgesia vol100 no 4 pp 1221ndash1222 2005

[37] C Rosow and P J Manberg ldquoBispectral index monitoringrdquoAnesthesiology Clinics of North America vol 19 no 4 pp 947ndash966 2001

[38] G J Jiang S-Z FanM FAbbod et al ldquoSample entropy analysisof EEG signals via artificial neural networks to model patientsrsquoconsciousness level based on anesthesiologists experiencerdquoBioMed Research International vol 2015 Article ID 343478 8pages 2015

[39] N Rehman and D P Mandic ldquoMultivariate empirical modedecompositionrdquo Proceedings of the Royal Society A Mathemat-ical Physical and Engineering Sciences vol 466 no 2117 pp1291ndash1302 2010

[40] N E Huang Z Shen S R Long et al ldquoThe empirical modedecomposition and the Hilbert spectrum for nonlinear andnon-stationary time series analysisrdquo Proceedings of the Royal


Society of London A Mathematical Physical and EngineeringSciences vol 454 no 1971 pp 903ndash995 1998

[41] P Flandrin G Rilling and P Goncalves ldquoEmpirical modedecomposition as a filter bankrdquo IEEE Signal Processing Lettersvol 11 no 2 pp 112ndash114 2004

[42] Z Wu and N E Huang ldquoEnsemble empirical mode decom-position a noise-assisted data analysis methodrdquo Advances inAdaptive Data Analysis vol 1 no 1 pp 1ndash41 2009

[43] J-R Huang S-Z Fan M F Abbod K-K Jen J-F Wuand J-S Shieh ldquoApplication of multivariate empirical modedecomposition and sample entropy in EEG signals via artificialneural networks for interpreting depth of anesthesiardquo Entropyvol 15 no 9 pp 3325ndash3339 2013

[44] J Bruhn H Ropcke B Rehberg T Bouillon and A HoeftldquoElectroencephalogram approximate entropy correctly classi-fies the occurrence of burst suppression pattern as increasinganesthetic drug effectrdquo Anesthesiology vol 93 no 4 pp 981ndash985 2000

[45] P Gifani H R Rabiee M H Hashemi P Taslimi and MGhanbari ldquoOptimal fractal-scaling analysis of human EEGdynamic for depth of anesthesia quantificationrdquo Journal of theFranklin Institute vol 344 no 3-4 pp 212ndash229 2007

[46] S Kumar A Kumar S Gombar A Trikha and S AnandldquoFractal dimension of electroencephalogram for assessmentof hypnosis state of patient during anaesthesiardquo InternationalJournal of Biomedical Engineering and Technology vol 10 no1 pp 30ndash37 2012

[47] K Hayashi N Mukai and T Sawa ldquoPoincare analysis of theelectroencephalogram during sevoflurane anesthesiardquo ClinicalNeurophysiology vol 126 no 2 pp 404ndash411 2015

[48] U Melia M Vallverdu X Borrat et al ldquoPrediction of noci-ceptive responses during sedation by linear and non-linearmeasures of EEG signals in high frequenciesrdquo PLoS ONE vol10 no 4 Article ID e0123464 2015

[49] R Ferenets T Lipping A Anier V Jantti S Melto and SHovilehto ldquoComparison of entropy and complexity measuresfor the assessment of depth of sedationrdquo IEEE Transactions onBiomedical Engineering vol 53 no 6 pp 1067ndash1077 2006

[50] J Schuttler and H Schwilden Modern Anesthetics SpringerBerlin Germany 2007

[51] U C Lee G A Mashour S Kim G-J Noh and B-M ChoildquoPropofol induction reduces the capacity for neural informationintegration implications for the mechanism of consciousnessand general anesthesiardquo Consciousness and Cognition vol 18no 1 pp 56ndash64 2009

[52] G Kumar C Stendall R Mistry K Gurusamy and D WalkerldquoA comparison of total intravenous anaesthesia using propofolwith sevoflurane or desflurane in ambulatory surgery system-atic review and meta-analysisrdquo Anaesthesia vol 69 no 10 pp1138ndash1150 2014

[53] H S Joo and W J Perks ldquoSevoflurane versus propofol foranesthetic induction a meta-analysisrdquo Anesthesia amp Analgesiavol 91 no 1 pp 213ndash219 2000

[54] E N Brown R Lydic and N D Schiff ldquoGeneral anesthesiasleep and comardquoTheNewEngland Journal ofMedicine vol 363no 27 pp 2638ndash2650 2010

Submit your manuscripts athttpwwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


[7 8] In particular the methods based on EEG for DOAevaluation have been developed rapidly The EEG signalswhich reflect the brainrsquos activities have been widely usedfor research and diagnosis especially for measuring theawareness level of patients EEG referring to brainrsquos electricalactivity is commonly recorded in a noninvasive approachwhich provides an available tool to study the human brainfor researchers and doctors [9] It has been widely usedfor measuring consciousness level of patients in medicalenvironment [10ndash12]

There are variousmethods based on EEG analysis appliedto monitor DOA recentlyThe bispectral index (BIS) monitorintroduced by Aspect Medical Systems Inc in 1994 [13ndash15] is widely used in the operation room for evaluatingthe DOA by analysis of EEG signals of patients duringsurgery BIS monitor has been proved as a reliable systemto measure the DOA except for several anaesthetic agentsin many researches [16 17] However the company thatintroduced the BIS monitor has not disclosed the detailedalgorithms In addition entropy monitors developed byDatex-Ohmeda produce response entropy (RE) and stateentropy (SE) to evaluate the irregularity in EEG signalsfor determining the DOA [18] The algorithm applied inthe Datex-Ohmeda entropy module calculates the RE andSE based on frequency domain approach called spectralentropy which is obtained by applying Shannon entropyto the power spectrum [19] However application of fastFourier transform (FFT) to estimate power spectrum maymiss the nonlinear and nonstationary properties of EEGsignals Although these two monitor systems are the mostpopular there are limitations Therefore an open sourceand time domain based method taking the nonlinear andnonstationary properties of EEG signals into considerationis need for monitoring DOA during surgery robustly andaccurately

The approximate entropy (ApEn) [20] and SampEn [21]algorithms are two powerful approaches proposed in appli-cation of determining the complexity of any time seriesAnd SampEn has been proved to perform better than ApEnfor monitoring DOA of patients during surgery in previousstudies [21ndash23] Nevertheless SampEn measures complexityof time series based on a single time scale so that it missesthe features associated with signal structure To overcomethis problem Costa et al introduced an improved methodnamedmultiscale entropy (MSE) to analyze the complexity ofbiological signals over multiple time scales [24 25] The EEGreflects the summation of human brainrsquos activity and containsthe information about neuronal dynamics underlying highand low frequency [26 27] Therefore MSE is appropriatefor obtaining the dynamics features related to multiple timescales and has been widely used in analysis of EEG recordings[28ndash30] Although MSE measure can explore the degree ofcomplexity over different time scales a single index whichindicates the DOA of patients during surgery is needed bysurgeons However we find that on the one hand manystudies applied MSE to distinguish the complexity of EEGthrough plotting entropy values over different time scalesoverall without considering time [31 32] or calculate entropyvalues on all time scales independently for monitoring DOA

[30] which is too complicated for surgeons to determinepatientsrsquo anesthesia level On the other hand in previousresearch [33] a single index was derived from entropy valuesbased on appropriate scales by averaging the scale dependententropies The limitation of this method is that entropiesrelated to each scale contribute unequally to measure thecomplexity and it is difficult to confirm the weights for eachindependent scale Therefore in this paper a new index isobtained from MSE analysis by combining the independententropies via ANN for measuring the DOA during surgery

TheANN is an extremely important and useful algorithmin machine learning inspired by biological neural networks[34] It can be used to adaptively and optimally estimatethe weights and functions which are generally unknown inadvance to depend on the input and target data by trainingvalidating and testing Therefore it has been widely used tosolve many tasks for classification and regression analysis inbiomedical engineering [35]

In this study simulated EEG corrupted with EOG andEEG collected from patients with different consciousnesslevel are analyzed by MSE to investigate the sensitivity toEOG and ability to distinguish the patientsrsquo states of entropiescorresponding to each independent scale Next we apply theMSE method to real EEG recordings collected from patientsduring surgery And then entropies and a gold standardare defined as input and target variables to train the ANNmodelThe outputs of ANN are severed as the new combinedindex for DOA monitoring BIS as a commercial index hasbeen approved by US Food and Drug Administration andmost widely used in operation room during surgery and ICUto monitor DOA although it is not perfect For exampleintraoperative awareness can occur during general anesthesiawith a small probability event even if BIS value is under60 according to recent researches [15 36 37] BIS is one ofthe technologies to accurately monitor the hypnotic effectsof general anesthetics and sedatives based on EEG signalsHowever the device is very expensive and the details of thealgorithms to calculate BIS index have not been disclosedSo it is necessary to create an open sources method forDOA monitoring accurately In this paper our aim is tocreate a new index which can accurately trace the changeof consciousness level of patients like BIS therefore BIS isused as a gold standard of DOA However it would be moreapplicable and reasonable if there is a real gold standard ofDOA as the target So in comparison ldquothe state of anestheticdepthrdquo called expert assessment of conscious level (EACL)[38] which is decided by five experienced anesthesiologistsbased on detailed recordings during surgery was used as thetarget to train ANN

EEG signals are always corrupted by artifacts such asEOG and EMG Generally the amplitude of EEG can beextremely less than artifacts so techniques are needed toremove EEG contaminants for accurate analysis In thispaper MEMD [39] based filter was used to remove artifactsfrom contaminated EEG signals On the one hand throughcomparison of performance of proposed method before andafter filtering we can indicate the robustness of proposedmethod to artifacts On the other hand MSE measurescomplexity of time series at different scales and filter can


2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(a)

2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(b)

2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(c)

2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(d)











119883 (119905) =

119873

sum

119894=1

119888

119894(119905) + 119903

119873(119905)

(1)


119894(119905) is the 119894th IMF and 119903



119883 (119905) =

119902

sum

119901

119888

119894(119905) (2)

where




10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(a)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(b)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(c)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(d)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(e)

Mean plusmn StdMean

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(f)



119883 (119905) =

119902

sum

1

119888

119894(119905) +

119873

sum

119901

119888

119894(119905) (3)



MEMD

EEG

White noise

White noise








template defined as

119883

119898(119894) = 119909

119894+1 119909

119894+2 119909

119894+3 119909

119894+119898minus1

1 ⩽ 119894 ⩽ 119873 minus 119898 + 1

(4)


119889 [119883

119898(119894) 119883

119898(119895)] = max (10038161003816

1003816

1003816

119883

119898(119894 + 119896) minus 119883

119898(119895 + 119896)

1003816

1003816

1003816

1003816

)

for 0 ⩽ 119896 ⩽ 119898 minus 1(5)


119883

119898(119894) denoted by 119861

119894and then

119861

119898

119894(119903) =

1

119873 minus 119898 minus 1

119861

119894

119861

119898(119903) =

1

119873 minus 119898

119873minus119898

sum

119894=1

119861

119898

119894(119903)

(6)


119860

119898(119903) =

1

119873 minus 119898

119873minus119898

sum

119894=1

119860

119898

119894(119903)

(7)


SampEn (119898 119903119873) = minus ln 119860119898(119903)

119861

119898(119903)

(8)


119910

(120591)

119895=

1

120591

119895120591

sum

119894=(119895minus1)120591+1

119883

119894 (9)






120591 = 1

120591 = 2

120591 = 3

120591 = i




y1 =x1 + x2

2y2 =

x3 + x42

y3 =x5 + x6

2yj =

xi + xi+12

y1 =x1 + x2 + x3

3y2 =

x4 + x5 + x63


3


i













05

1

15

2

25

Sam

pEn





3 Results






N

ANN

Network outputs

Input data (MSE)

Target data

Scale 1

Scale 2

Scale 3

Scale N

Output layer

20 1


050100

0

1

2

0

1

2

0

1

2

0

1

2

050100

(a)

+ f

w1

w2

w3

wn

Input 1

Input 2

Input 3

Input n b

Output

(b)



0

01

02

03

04

CV






Stage 1Stage 3


0

05

1

15

2

25En

tropy

val

ues

(a)

Stage 1Stage 3


0

05

1

15

2

25

Entro

py v

alue

s(b)










0

02

04

06

08

1C

orre

latio

n co

effici

ents

(Cor

rcoe

f)


(a)


0

02

04

06

08

1

Cor

relat

ion

coeffi

cien

ts (C

orrc

oef)


(b)















4 Discussion
































5 Conclusions




Acknowledgments


References
































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(a)

2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(b)

2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(c)

2 4 6 8 100Time (s)

Am

plitu

de (120583

v)

minus1000

0

1000

(d)











119883 (119905) =

119873

sum

119894=1

119888

119894(119905) + 119903

119873(119905)

(1)


119894(119905) is the 119894th IMF and 119903



119883 (119905) =

119902

sum

119901

119888

119894(119905) (2)

where




10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(a)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(b)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(c)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(d)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(e)

Mean plusmn StdMean

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(f)



119883 (119905) =

119902

sum

1

119888

119894(119905) +

119873

sum

119901

119888

119894(119905) (3)



MEMD

EEG

White noise

White noise








template defined as

119883

119898(119894) = 119909

119894+1 119909

119894+2 119909

119894+3 119909

119894+119898minus1

1 ⩽ 119894 ⩽ 119873 minus 119898 + 1

(4)


119889 [119883

119898(119894) 119883

119898(119895)] = max (10038161003816

1003816

1003816

119883

119898(119894 + 119896) minus 119883

119898(119895 + 119896)

1003816

1003816

1003816

1003816

)

for 0 ⩽ 119896 ⩽ 119898 minus 1(5)


119883

119898(119894) denoted by 119861

119894and then

119861

119898

119894(119903) =

1

119873 minus 119898 minus 1

119861

119894

119861

119898(119903) =

1

119873 minus 119898

119873minus119898

sum

119894=1

119861

119898

119894(119903)

(6)


119860

119898(119903) =

1

119873 minus 119898

119873minus119898

sum

119894=1

119860

119898

119894(119903)

(7)


SampEn (119898 119903119873) = minus ln 119860119898(119903)

119861

119898(119903)

(8)


119910

(120591)

119895=

1

120591

119895120591

sum

119894=(119895minus1)120591+1

119883

119894 (9)






120591 = 1

120591 = 2

120591 = 3

120591 = i




y1 =x1 + x2

2y2 =

x3 + x42

y3 =x5 + x6

2yj =

xi + xi+12

y1 =x1 + x2 + x3

3y2 =

x4 + x5 + x63


3


i













05

1

15

2

25

Sam

pEn





3 Results






N

ANN

Network outputs

Input data (MSE)

Target data

Scale 1

Scale 2

Scale 3

Scale N

Output layer

20 1


050100

0

1

2

0

1

2

0

1

2

0

1

2

050100

(a)

+ f

w1

w2

w3

wn

Input 1

Input 2

Input 3

Input n b

Output

(b)



0

01

02

03

04

CV






Stage 1Stage 3


0

05

1

15

2

25En

tropy

val

ues

(a)

Stage 1Stage 3


0

05

1

15

2

25

Entro

py v

alue

s(b)










0

02

04

06

08

1C

orre

latio

n co

effici

ents

(Cor

rcoe

f)


(a)


0

02

04

06

08

1

Cor

relat

ion

coeffi

cien

ts (C

orrc

oef)


(b)















4 Discussion
































5 Conclusions




Acknowledgments


References
































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



















119883 (119905) =

119873

sum

119894=1

119888

119894(119905) + 119903

119873(119905)

(1)


119894(119905) is the 119894th IMF and 119903



119883 (119905) =

119902

sum

119901

119888

119894(119905) (2)

where




10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(a)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(b)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(c)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(d)

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(e)

Mean plusmn StdMean

10 20 30 40 50 60 70 80 90 1000Time (min)

0

50

100

EACL

(f)



119883 (119905) =

119902

sum

1

119888

119894(119905) +

119873

sum

119901

119888

119894(119905) (3)



MEMD

EEG

White noise

White noise








template defined as

119883

119898(119894) = 119909

119894+1 119909

119894+2 119909

119894+3 119909

119894+119898minus1

1 ⩽ 119894 ⩽ 119873 minus 119898 + 1

(4)


119889 [119883

119898(119894) 119883

119898(119895)] = max (10038161003816

1003816

1003816

119883

119898(119894 + 119896) minus 119883

119898(119895 + 119896)

1003816

1003816

1003816

1003816

)

for 0 ⩽ 119896 ⩽ 119898 minus 1(5)


119883

119898(119894) denoted by 119861

119894and then

119861

119898

119894(119903) =

1

119873 minus 119898 minus 1

119861

119894

119861

119898(119903) =

1

119873 minus 119898

119873minus119898

sum

119894=1

119861

119898

119894(119903)

(6)


119860

119898(119903) =

1

119873 minus 119898

119873minus119898

sum

119894=1

119860

119898

119894(119903)

(7)


SampEn (119898 119903119873) = minus ln 119860119898(119903)

119861

119898(119903)

(8)


119910

(120591)

119895=

1

120591

119895120591

sum

119894=(119895minus1)120591+1

119883

119894 (9)






120591 = 1

120591 = 2

120591 = 3

120591 = i




y1 =x1 + x2

2y2 =

x3 + x42

y3 =x5 + x6

2yj =

xi + xi+12

y1 =x1 + x2 + x3

3y2 =

x4 + x5 + x63


3


i













05

1

15

2

25

Sam

pEn





3 Results






N

ANN

Network outputs

Input data (MSE)

Target data

Scale 1

Scale 2

Scale 3

Scale N

Output layer

20 1


050100

0

1

2

0

1

2

0

1

2

0

1

2

050100

(a)

+ f

w1

w2

w3

wn

Input 1

Input 2

Input 3

Input n b

Output

(b)



0

01

02

03

04

CV






Stage 1Stage 3


0

05

1

15

2

25En

tropy

val

ues

(a)

Stage 1Stage 3


0

05

1

15

2

25

Entro

py v

alue

s(b)










0

02

04

06

08

1C

orre

latio

n co

effici

ents

(Cor

rcoe

f)


(a)


0

02

04

06

08

1

Cor

relat

ion

coeffi

cien

ts (C

orrc

oef)


(b)















4 Discussion
































5 Conclusions




Acknowledgments


References
































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















MEMD

EEG

White noise

White noise








template defined as

119883

119898(119894) = 119909

119894+1 119909

119894+2 119909

119894+3 119909

119894+119898minus1

1 ⩽ 119894 ⩽ 119873 minus 119898 + 1

(4)


119889 [119883

119898(119894) 119883

119898(119895)] = max (10038161003816

1003816

1003816

119883

119898(119894 + 119896) minus 119883

119898(119895 + 119896)

1003816

1003816

1003816

1003816

)

for 0 ⩽ 119896 ⩽ 119898 minus 1(5)


119883

119898(119894) denoted by 119861

119894and then

119861

119898

119894(119903) =

1

119873 minus 119898 minus 1

119861

119894

119861

119898(119903) =

1

119873 minus 119898

119873minus119898

sum

119894=1

119861

119898

119894(119903)

(6)


119860

119898(119903) =

1

119873 minus 119898

119873minus119898

sum

119894=1

119860

119898

119894(119903)

(7)


SampEn (119898 119903119873) = minus ln 119860119898(119903)

119861

119898(119903)

(8)


119910

(120591)

119895=

1

120591

119895120591

sum

119894=(119895minus1)120591+1

119883

119894 (9)






120591 = 1

120591 = 2

120591 = 3

120591 = i




y1 =x1 + x2

2y2 =

x3 + x42

y3 =x5 + x6

2yj =

xi + xi+12

y1 =x1 + x2 + x3

3y2 =

x4 + x5 + x63


3


i













05

1

15

2

25

Sam

pEn





3 Results






N

ANN

Network outputs

Input data (MSE)

Target data

Scale 1

Scale 2

Scale 3

Scale N

Output layer

20 1


050100

0

1

2

0

1

2

0

1

2

0

1

2

050100

(a)

+ f

w1

w2

w3

wn

Input 1

Input 2

Input 3

Input n b

Output

(b)



0

01

02

03

04

CV






Stage 1Stage 3


0

05

1

15

2

25En

tropy

val

ues

(a)

Stage 1Stage 3


0

05

1

15

2

25

Entro

py v

alue

s(b)










0

02

04

06

08

1C

orre

latio

n co

effici

ents

(Cor

rcoe

f)


(a)


0

02

04

06

08

1

Cor

relat

ion

coeffi

cien

ts (C

orrc

oef)


(b)















4 Discussion
































5 Conclusions




Acknowledgments


References
































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















120591 = 1

120591 = 2

120591 = 3

120591 = i




y1 =x1 + x2

2y2 =

x3 + x42

y3 =x5 + x6

2yj =

xi + xi+12

y1 =x1 + x2 + x3

3y2 =

x4 + x5 + x63


3


i













05

1

15

2

25

Sam

pEn





3 Results






N

ANN

Network outputs

Input data (MSE)

Target data

Scale 1

Scale 2

Scale 3

Scale N

Output layer

20 1


050100

0

1

2

0

1

2

0

1

2

0

1

2

050100

(a)

+ f

w1

w2

w3

wn

Input 1

Input 2

Input 3

Input n b

Output

(b)



0

01

02

03

04

CV






Stage 1Stage 3


0

05

1

15

2

25En

tropy

val

ues

(a)

Stage 1Stage 3


0

05

1

15

2

25

Entro

py v

alue

s(b)










0

02

04

06

08

1C

orre

latio

n co

effici

ents

(Cor

rcoe

f)


(a)


0

02

04

06

08

1

Cor

relat

ion

coeffi

cien

ts (C

orrc

oef)


(b)















4 Discussion
































5 Conclusions




Acknowledgments


References
































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















N

ANN

Network outputs

Input data (MSE)

Target data

Scale 1

Scale 2

Scale 3

Scale N

Output layer

20 1


050100

0

1

2

0

1

2

0

1

2

0

1

2

050100

(a)

+ f

w1

w2

w3

wn

Input 1

Input 2

Input 3

Input n b

Output

(b)



0

01

02

03

04

CV






Stage 1Stage 3


0

05

1

15

2

25En

tropy

val

ues

(a)

Stage 1Stage 3


0

05

1

15

2

25

Entro

py v

alue

s(b)










0

02

04

06

08

1C

orre

latio

n co

effici

ents

(Cor

rcoe

f)


(a)


0

02

04

06

08

1

Cor

relat

ion

coeffi

cien

ts (C

orrc

oef)


(b)















4 Discussion
































5 Conclusions




Acknowledgments


References
































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Stage 1Stage 3


0

05

1

15

2

25En

tropy

val

ues

(a)

Stage 1Stage 3


0

05

1

15

2

25

Entro

py v

alue

s(b)










0

02

04

06

08

1C

orre

latio

n co

effici

ents

(Cor

rcoe

f)


(a)


0

02

04

06

08

1

Cor

relat

ion

coeffi

cien

ts (C

orrc

oef)


(b)















4 Discussion
































5 Conclusions




Acknowledgments


References
































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















0

02

04

06

08

1C

orre

latio

n co

effici

ents

(Cor

rcoe

f)


(a)


0

02

04

06

08

1

Cor

relat

ion

coeffi

cien

ts (C

orrc

oef)


(b)















4 Discussion
































5 Conclusions




Acknowledgments


References
































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
























4 Discussion
































5 Conclusions




Acknowledgments


References
































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of














































5 Conclusions




Acknowledgments


References
































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of










































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















eeg signals analysis using multiscale entropy for depth of

Documents