Classification of EEG Signals Based on Pattern Recognition Approach

12 Pages·2017·2.18 MB·English

METHODS
published:21November2017
doi:10.3389/fncom.2017.00103
Classification of EEG Signals Based
on Pattern Recognition Approach
HafeezUllahAmin*,WajidMumtaz,AhmadRaufSubhani,
MohamadNaufalMohamadSaadandAamirSaeedMalik*
CentreforIntelligentSignalandImagingResearch(CISIR),DepartmentofElectricalandElectronicEngineering,Universiti
TeknologiPetronas,SeriIskandar,Malaysia
Feature extraction is an important step in the process of electroencephalogram
(EEG) signal classification. The authors propose a “pattern recognition” approach
that discriminates EEG signals recorded during different cognitive conditions. Wavelet
based feature extraction such as, multi-resolution decompositions into detailed and
approximate coefficients as well as relative wavelet energy were computed. Extracted
relative wavelet energy features were normalized to zero mean and unit variance and
thenoptimizedusingFisher’sdiscriminantratio(FDR)andprincipalcomponentanalysis
(PCA). A high density EEG dataset validated the proposed method (128-channels)
by identifying two classifications: (1) EEG signals recorded during complex cognitive
tasksusingRaven’sAdvanceProgressiveMetric(RAPM)test;(2)EEGsignalsrecorded
during a baseline task (eyes open). Classifiers such as, K-nearest neighbors (KNN),
SupportVectorMachine(SVM),Multi-layerPerceptron(MLP),andNaïveBayes(NB)were
then employed. Outcomes yielded 99.11% accuracy via SVM classifier for coefficient
approximations (A5) of low frequencies ranging from 0 to 3.90Hz. Accuracy rates for
Editedby: detailed coefficients were 98.57 and 98.39% for SVM and KNN, respectively; and for
DanChen, detailed coefficients (D5) deriving from the sub-band range (3.90–7.81Hz). Accuracy
WuhanUniversity,China
ratesforMLPandNBclassifierswerecomparableat97.11–89.63%and91.60–81.07%
Reviewedby:
for A5 and D5 coefficients, respectively. In addition, the proposed approach was
DuanLi,
UniversityofMichigan,UnitedStates also applied on public dataset for classification of two cognitive tasks and achieved
JamieSleigh,
comparable classification results, i.e., 93.33% accuracy with KNN. The proposed
UniversityofAuckland,NewZealand
scheme yielded significantly higher classification performances using machine learning
*Correspondence:
HafeezUllahAmin classifiers compared to extant quantitative feature extraction. These results suggest
[email protected] the proposed feature extraction method reliably classifies EEG signals recorded during
AamirSaeedMalik
cognitivetaskswithahigherdegreeofaccuracy.
[email protected]
Keywords:featureextraction,featureselection,machinelearningclassifiers,electroencephalogram(EEG)
Received:02May2017
Accepted:01November2017
Published:21November2017
INTRODUCTION
Citation:
AminHU,MumtazW,SubhaniAR, Clinicians use the electroencephalogram (EEG) as a standard neuroimaging tool for the study
SaadMNMandMalikAS(2017)
of neuronal dynamics within the human brain. Data extracted from EEG results reflect the
ClassificationofEEGSignalsBased
process of an individual’s information processing (Grabner et al., 2006). Recent technological
onPatternRecognitionApproach.
Front.Comput.Neurosci.11:103. advanceshaveincreasedthescopeofEEGrecordingabilitiesbyusingdensegroupsofelectrodes
doi:10.3389/fncom.2017.00103 includingarraysof128,256,and512electrodesattachedtothecranium.Visualinspectionofthese
FrontiersinComputationalNeuroscience|www.frontiersin.org 1 November2017|Volume11|Article103
Aminetal. EEGSignalsAnalysis
massivedatasetsiscumbersome(Übeyli,2009)forexistingEEG- methods including Fisher’s discriminant ratio (FDR) and
basedanalysistechniques.Hence,optimizedfeatureextractionof principalcomponentanalysis(PCA)toeliminatenon-significant
relevantEEGdataisessentialtoimprovethequalityofcognitive features, which, in turn, optimized the “features” data set. As
performance evaluations, especially since it directly impacts a for classification, we employed K-nearest neighbors (KNN),
classifier’s performance (Iscan et al., 2011). In addition, more SupportVectorMachine(SVM),Multi-layerPerceptron(MLP),
expressive features enhance classification performance; hence, andNaïveBayes(NB)tooptimizethe“features”setevenfurther
featureextractionhasbecomethemostcriticallysignificantstep forourpurposeofEEGpatternclassification.Resultswerethen
inEEGdataclassification. comparedwithexistingquantitativemethodstoconfirmrather
Severalresearchersinvestigated“Quantitative-EEG”(QEEG) robust outcomes. Two EEG datasets were used to validate the
fortheevaluationofneuralactivityduringcognitivetasks.They proposed method. Dataset I comprise complex cognitive task
usedtimeandfrequencydomainfeaturessuchas,entropy,power (class 1) and baseline eyes open (class 2) task; while dataset II
spectrum, autoregressive coefficients, and individual frequency comprises two cognitive tasks: mental multiplication (class 1)
bands, etc. (Doppelmayr et al., 2005; Zarjam et al., 2012). andmentallettercomposing(class2).
Frequency characteristics depend on neuronal activity and are Section Materials and Methods describes materials and
groupedintoseveralbands(delta,theta,alpha,beta,andgamma; methods, followed by section Experimental Results and
Amin and Malik, 2013) that have been linked to cognitive Discussion, which presents results and discussion, followed by
processes. sectionLimitationsandconcludingremarks.
Severalfeatureextractionmethodshavebeenreportedinthe
literature.Theseincludetime-frequencydomainandthewavelet MATERIALS AND METHODS
transform(WT)(Iscanetal.,2011).WT-basedanalysisishighly
effectivefornon-stationaryEEGsignalscomparedtotheshort- This section provides details of experimental tasks and the
time Fourier transformation (STFT). Moreover, wavelet-based dataset used in this study. In addition, we briefly describe
features,includingwaveletentropy(Rossoetal.,2001),wavelet the classification algorithms employed and the DWT, as
coefficients (Orhan et al., 2011), and wavelet statistical features well as the computations used for wavelet and relative
(mean, median, and standard deviations) have been reported energy.
for the evaluation of normal EEG patterns and for clinical
Raven’s Advance Progressive Matric Test
applications (Yazdani et al., 2009; Garry et al., 2013). However,
significant gaps in the literature exist regarding cognitive load (RAPM)
studies and approaches to pattern recognition. Many studies Raven’s Advance Progressive Matric Test (Raven, 2000) is
employedmultiplecognitivetaskssuchas,multiplication,mental a non-verbal tool that measures levels of an individual’s
object rotation, mental letter composing, and visual counting intellectual ability. It is commonly used to explicitly measure
(Xue et al., 2003). These tasks are simple in nature and may twocomponentsofgeneralcognitiveability(Raven,2000):“the
not induce a high enough load to activate correlative cognitive ability to draw meaning out of confusion, and the ability to
neuronalnetworksthatgeneratedetectableelectricalpotentials. recallandreproduceinformationthathasbeenmadeexplicitand
Furthermore, from a pattern recognition perspective, several communicatedfromonetoanother.”TheRAPMtestcomprises
studiesexcluded“featurenormalization”and“featureselection” 48 patterns (questions) divided into two sets (I and II). Set-
steps (Xue et al., 2003; Daud and Yunus, 2004), while many I contains 12 practice patterns and Set-II contains 36 patterns
others used very few instances (observations) as input for used to assess cognitive ability. As shown in Figure1, each
classifiers that query the output of classification algorithms patterncontainsa3 3-cellstructurewhereeachcellcontainsa
×
(Lin and Hsieh, 2009; Guo et al., 2011). In addition, some geometricalshape,exceptfortheemptyright-bottomcell.Eight
studies failed to cross validate their proposed classification multipleoptionsarepresentedassolutionsfortheemptycell.A
procedures. scoreof“1”isassignedforeachcorrectanswerandascoreof“0”
The present study utilized “Raven’s Advance Progressive foranincorrectanswer.Theprocessingtimeis10and40minfor
Metric” (RAPM)—commonly used to measure inter-individual setsIandII,respectively.
differences in cognitive performance—as its standard
psychometric cognitive task to stimulate EEG data (Raven, Description of Dataset I
2000; Neubauer and Fink, 2009). RAPM requires higher ProceduresadoptedfortherecordingandpreprocessingofEEG
cognitiveprocessingresourcesandreasoningabilitytoperform dataarefoundinourpreviousstudies[7,8].Afterpreprocessing,
its tasks, which are non-verbal psychometric tests that require wedividedthedatasetintotwoclassesasfollows:Class1included
inductive reasoning considered an indicator of cognitive EEGdatafromeightexperimentalsubjectsrecordedduringtheir
performance (Raven, 2000). The authors propose a “pattern RAPM task performances; Class 2 included EEG data recorded
recognition” approach comprising feature extraction, feature witheyes-open(EO)fromthesameeightsubjects’.Asmentioned
normalization,featureselection,featureclassification,andcross intheRAPMtaskdescription,eachsubjectcompleted36patterns
validation (Figure5). Wavelet coefficients were extracted using within a specified period. Hence, these EEG recordings were
the discrete wavelet transform (DWT) as well as relative sub- time-markedfortheonsetofRAPMpatterndisplay,andagain
bandenergies,whichwerethenstandardizedtozeromeanand at the end of a subject’s response, specifically when pressing a
unit variance. The feature selection process utilized statistical button indicating a solution. Each subject’s EEG recording was
FrontiersinComputationalNeuroscience|www.frontiersin.org 2 November2017|Volume11|Article103
Aminetal. EEGSignalsAnalysis
MentalLetterCompositionTask
Inthistask,theparticipantswere askedto mentallycompose a
lettertoafriendorrelativewithoutvocalizing.
The Discrete Wavelet Transform (DWT)
Discrete Wavelet Transform decomposition includes successive
high and low pass filtering of a time series with a down-
sampling rate of 2. The high pass filter [g(n)] is the discrete
“mother” wavelet and the low pass filter [h(n)] is its mirror
version (Subasi, 2007). The “mother” wavelet [Daubechies
wavelet(db4)]andcorrespondingscalingfunctionareshownin
Figure2.
Outputs from initial high pass and low pass filters are
called “approximations” and “detailed” coefficients (A1
and D1), respectively. A1 is disintegrated further and the
procedure repeated until reaching a specified number of
decomposition levels (see Figure3; Jahankhani et al., 2006;
Subasi,2007).
Thescaling[ϕ (n)]andwaveletfunctions[ψ (n)]both
j,k j,k
dependonlowpassandhighpassfilters,respectively.Theseare
denotedasfollows:
φj,k(n) 2−j/2h 2−jn k (1)
= (cid:0) − (cid:1)
FIGURE1|AsimpleRaven’sstylepattern(Aminetal.,2015b). ψj,k(n) 2−j/2g 2−jn k (2)
= (cid:0) − (cid:1)
Wheren 0, 1, 2, ..., M 1 j 0, 1, 2,..., J 1 k
= − ; = − ; =
0, 1, 2, ..., 2j 1 J 5 and
segmentedaccordingtothenumberofpatternssolved.EachEEG − ; = ;
Misthelengthof thesignal(GonzalezandWoods,2002).
segment was also considered an observation; thus, producing
Approximation coefficients (A) and detailed coefficients
36 observations per subject. However, a few patterns went i
(D )attheithlevelaredeterminedasfollows(Orhanetal.,2011):
unanswered (unsolved) by some subjects, which EEG segments i
wereexcludedfromtheanalysis.Weobserved280observations
1
forClass1.Tomaintainabalancebetweenclasses,EOandEC A x(n).φ (n) (3)
i = √MX j,k
(eyes-closed)dataforeachsubjectweresegmentedaccordingto n
the number of attempted RAPM patterns. As a result, EO data 1
D x(n).ψ (n) (4)
alsoincluded280observationsforClass2. i = √MX j,k
n
Description of Dataset II Relative and Total Wavelet Sub-band
EEGdatasetutilizedinthisstudywasoriginallyreportedbyKeirn Energy
and Aunon (1990) and publically available for reuse. The EEG Wavelet energy for each decomposition level (i 1,...,l) is
datawasrecordedbyplacingelectrodesoverC3,C4,P3,P4,O1, determinedasfollows: =
andO2positionsaccordingto10–20montageandreferencedto
linkedmastoids,A1andA2.Theimpedancesofalltheelectrodes N
2
were kept below 5 k(cid:127). The data were digitized at 250Hz with EDi =X(cid:12)Dij(cid:12) , i = 1, 2, 3,...,l (5)
a Lab Master 12-bit A/D converter mounted on a computer. j 1(cid:12) (cid:12)
=
Seven participants, 21–48 years old, recorded EEG data during
cognitive tasks. The data was recorded for duration of 10s in Wherel 5,reflectsthelevelofdecompositions
=
eachtrialandeachtaskwasrepeatedfivetimes(formoredetail
N
of dataset, see Keirn and Aunon, 1990 original work). For this 2
E A , i l (6)
study, we tested our proposed approach on two cognitive tasks Ai =X(cid:12) ij(cid:12) =
j 1(cid:12) (cid:12)
employedfromKeirnandAunon’sdatasetanddescribedbelow. =
Therefore,fromEquations(5)and(6),totalenergycanbedefined
MentalMultiplicationTask
as:
Theparticipantsweregivennontrivialmultiplicationproblems,
such as, multiply two multi-digit numbers, and were asked to l
solve them without vocalizing or making any other physical E  E E  (7)
Total = X Di + Al
movements. i 1 
=
FrontiersinComputationalNeuroscience|www.frontiersin.org 3 November2017|Volume11|Article103
Aminetal. EEGSignalsAnalysis
FIGURE2|Motherwaveletandscalingfunction(db4).
FIGURE3|DWTsub-banddecomposition.
The normalized energy values represent relative wavelet energy SupportVectorMachine(SVM)
(see Figure4 for an example of total and relative sub-band The SVM is a supervised learning algorithm that uses a kernel
energy). trick to transform input data into higher dimensional space,
afterwhichitsegregatesthedataviaahyper-planwithmaximal
margins.Duetoitsabilitytomanagelargedatasets,thealgorithm
E
j
Er , (8) is widely used for binary classification problems in machine
= E
Total learning.FormoredetailsonSVM,see(Hsuetal.,2003).
MultilayerPerceptron(MLP)
WhereE E orE .
j = Di=1,...,5 Ai=5 MLP is a non-linear neural network based method comprising
three sequential layers: input, hidden and output, respectively,
Classification Algorithms
where the hidden layer transmits input data to the output
Machine learning classifiers used in this study are now briefly layer. However, the MLP model can cause over-fitting due to
described. insufficientorexcessivenumbersofneurons.Forourpurposes,
Aclassifierutilizesvaluesforindependentvariables(features) weemployedtheMLPmodelwithfivehiddenneurons.
as input to predict the corresponding class to which an
independent variable belongs (Pereira et al., 2009). A classifier NaïveBayes(NB)
hasanumberofparametersthatrequiretrainingfromatraining The NB classifier provides simple and efficient probabilistic
dataset. A trained classifier will model the association between classification based on Bayes’ theorem, which posits that
classesandcorrespondingfeaturesandiscapableofidentifying extracted features are not dependent. The NB model uses (i) a
new instances in an unseen testing dataset. We employed the maximumprobabilityalgorithmtodeterminetheclassofearlier
followingclassificationmethodstodemonstratetheeffectiveness probabilities, and (ii) a feature’s probability distribution from a
ofthisstudy’sproposedtechnique. training dataset. Results are then employed with a maximized
FrontiersinComputationalNeuroscience|www.frontiersin.org 4 November2017|Volume11|Article103
Aminetal. EEGSignalsAnalysis
FIGURE4|EEGsignalenergyandrelativesub-bandenergy.
posteriori decision treeto findthespecificclasslabelfor a new for either training or testing, where each instance is employed
testinstance(Hanetal.,2011). for validation exactly once. For our purposes, we used 10-
fold cross-validation to train and test extracted features for all
k-NearestNeighbor(k-NN) classifiers.
The k-nearest neighbor is a supervised learning algorithm that
identifiesatestingsample’sclassaccordingthemajorityclassof The Proposed Scheme
k-nearesttrainingsamples;i.e.,aclasslabelisallocatedtoanew Figure5 summarizes the proposed feature extraction scheme,
instanceofthemostcommonclassamongstKNNinthe“feature” whichcomprisesthefollowingsteps:
space.Inthisstudy,kvaluewassettothree.See(Pereiraetal., Steps:
2009;Hanetal.,2011)fordetailsonk-NN,SVM,MLP,andNB
a. FeatureExtraction
machinelearningclassifiers.
1. DecompositionofEEGsignalintosub-bandsusingDWT
K-foldCrossValidation 2. Computationofeachsub-band’srelativeenergy
All classification models in the present work were trained and
Repeat steps 1 and 2 for all channels for each subject and for
tested with EEG data and then confirmed using k-fold cross
eachsegment(question)
validation, whichis a commonly used technique thatcompares
b. FeatureVisualizationandStandardization
(i)performancesoftwoclassificationalgorithms,or(ii)evaluates
theperformanceofasingleclassifieronagivendataset(Wong, 1. Standardizationofextractedfeaturestozeromeanandunit
2015). It has the advantage of using all instances in a dataset variance
FrontiersinComputationalNeuroscience|www.frontiersin.org 5 November2017|Volume11|Article103
Aminetal. EEGSignalsAnalysis
FIGURE5|ProposedschemeforfeatureextractionandclassificationofEEGsignals.
c. FeatureSelection RelativeEnergyFeatureMatrix(−→Fr) [ErA5(280 128)]
= ×
1. ApplicationofFDRtothe“features”setfollowedbysorting
Where,thenumberofchannelsis128;thenumberofinstances
features in descending order according to the power of
in each class is 280; E represents relative energy at
discrimination rA5
approximation coefficients A5 (0–3.90Hz). Similarly, feature
2. SelectionofsubsetfeaturesabovetheFDRmedianofsorted
matricesforD –D coefficientswereidenticallyrepresentedfor
featuresaslistedinstep1 2 5
eachclassandeachexperimentalsubject.
3. Transforming selected subset features to principal
components followed by sorting in descending order FeatureVisualizationandNormalization
accordingtoprincipalvalue
Featurevisualizationisanimportantstepbeforetheapplication
4. Selectionofprincipalcomponentsupto95%variance
of any normalization method. Features should be visualized to
5. Reconstructionofcorrespondingfeaturevectors
checkthedistributionoffeaturevalues(Figure7).Inthisstudy,
Repeatsteps1to5forallsub-bandsinthe“features”set featureswerestandardizedasfollows:
d. FeatureClassification
x µ
x − (9)
1. ClassificationofselectedfeaturesetsviaKNN,SVM,MLP ´ = σ
andNB
Wherexistheoriginalfeaturevalue;µisthemean;andσ isthe
2. Evaluation of classifier performances with 10-cross
standarddeviationofthe“features”set,respectively,andx´isthe
validation and assessing degrees of accuracy, sensitivity
normalizedfeaturevalue.
andspecificity
Repeatsteps1and2foreachsub-bandofthe“features”set FeatureSelection
The main objective of the feature selection step in pattern
FeatureExtraction recognition is to select a subset from large numbers of
The EEG signal was decomposed into sub-band frequencies by available features that more robustly discriminate for purposes
using the discrete DWT with Daubechies 4 Wavelet to level ofclassification(Theodoridisetal.,2010).Inthisstudy,FDRand
5. Approximate and detailed coefficients were then computed PCAwereusedtooptimizefeaturesselection.
(see Figure6 for example). Table1 presents one channel’s sub- FDR is an independent type of class distribution and a
band relative energy percentage and frequency range for a quantifier of the discriminating power of individual features
singleexperimentalsubject.Totalandrelativesub-bandenergies betweenclasses.Ifm andm aremeanvalues,andσ2andσ2are
1 2 1 2
were then computed from extracted wavelet coefficients. We respectivevariancesforafeatureinbothclasses,FDRisdefined
calculated relative wavelet energy(ErD1,ErD2,...,ErA5) by using as:
Equation(8).
We computed relative energy features for all experimental (m1 m2)
FDR − (10)
subjectsandforalldatacollectedfromallchannels.Accordingly, = σ2 σ2
(cid:0) 1 − 2(cid:1)
for dataset I, the feature matrix describing relative energy for
a single subject in each EEG condition and for each sub-band PCA(AbdiandWilliams,2010)selectsformutuallyuncorrelated
(detailedorapproximated)becomes: features; hence, it avoids redundancy in the feature set. In a
FrontiersinComputationalNeuroscience|www.frontiersin.org 6 November2017|Volume11|Article103
Aminetal. EEGSignalsAnalysis
FIGURE6|A5andD1–D5componentsofanexperimentalsubject’sEEGsignalduringacognitivetask.
TABLE1|FrontallobeF3channels. Step2:Singularvaluedecomposition(SVD)isappliedto(S)as
wellasto(l)singularvaluesandsingularvectors(λ)and(aǫRl);
i i
Levels Waveletenergy% Waveletcoefficients Frequencybands(Hz)
then(i 1,2,...,l)arecomputed.Singularvaluesareorganized
=
in descending order (λ λ λ), after which correct
1 0.66 D1 62.50–125 1 ≥ 2 ≥ ··· ≥ l
pairing of singular values with corresponding singular vectors
2 2.31 D2 31.25–62.50
areensured.Inaddition,(m)largestsingularvaluesareselected.
3 7.63 D3 15.62–31.25
Normally,(m)isselectedtoincludeacertainpercentageoftotal
4 10.77 D4 7.81–15.62
energy(95%).Singularvalues(λ λ λ )arecalled(m)
5 21.55 D5 3.90–7.81 1≥ 2≥···≥ m
principalcomponents.Respective(column)singularvectors(a,i
5 57.12 A5 0–3.90 i
1,2,...,m)arethenusedtoconstructthetransformationmatrix:
=
Relativeenergysub-bandpercentagesandfrequencyrange.
A [a a a ... a ] (13)
1 2 3 m
=
featureset(X),containing(l)featuresof(N)exampleseach,such
that(xi ǫ Rl,i 1,2,...,N),PCAaimstolinearlytransformthe Each l-dimensional vector in the original space (X), is
=
feature set and thus obtain a new set of samples (y), in which transformedtoanm-dimensionalvector(y),viatransformation
componentsof(y)areuncorrelatedperEquation(12). (y ATx).Inotherwords,theithelementof(y),i.e.[y(i)],isthe
=
projectionof(x)on[a (y(i) ATx)].
y ATx (11) i = i
=
FeatureClassification
Where(A)isatransformationmatrixorarrangementofsingular Theoptimized“features”setwasvisualizedbyusingprobability
vectorsthatcorrespondtosignificantsingularvalues. distribution functions (PDFs) and the ROC curve to check for
ThePCAprocessrequiresanumberofsteps: any overlapping in each selected feature for both classes (see
Step 1: Covariance matrix (S) of feature matrix (X) is Figure8: EO vs. RAPM). Each selected feature yields a partial
estimated: overlap and ROC values that are >0.9 or close to 1.0, which
confirm the discriminating power of the selected “features” set.
N
S xxT (12) Finally,theclassifiers,KNN,SVM,MLP,andNB,wereemployed
=X i i forthediscriminationofbothclasses.
i 1
=
FrontiersinComputationalNeuroscience|www.frontiersin.org 7 November2017|Volume11|Article103
Aminetal. EEGSignalsAnalysis
FIGURE7|Featurevisualizationofonesamplewithinadatasetusingnormalprobabilityandhistogram.
EXPERIMENTAL RESULTS AND composingtask.Theextractedfeatureswerereducedtooptimum
DISCUSSION numberoffeaturesbyusingFDRandPCA.Wefurtherapplied
machine learning algorithms, i.e., SVM with RBF kernel, MLP
Experimental Setup
with five hidden layers, KNN with k 3, and NB to classify
According to the 10-fold cross validation process, we divided the extracted features to all decompos=ition levels (D2–D5 and
the dataset into 10 subsets of equal instances with nine subsets A5). The detail coefficient D1 reflects high frequency (62.5–
employed for classifier training and one subset for classifier 125Hz)components,thusconsideredasnoiseandexcludedfrom
testing.Theprocesswasrepeated10timessothateachsubsetwas classification.
testedforclassification.Classifierperformancewasmeasuredfor
accuracy,sensitivity,specificity,precision,andtheKappastatistic
ResultsofDatasetI
(Aminetal.,2015a),eachdefinedasfollows:
The performance of SVM and KNN classifiers showed 99.11
Totalno.ofcorrectlyclassifiedinstances and 98.21% accuracies each for A5 approximation coefficients,
Accuracy 100 and 98.57 and 98.39% accuracies for D5 details coefficients,
= Totalnumbersofinstances ×
(14) respectively, as shown in Tables2, 3. This impressive
TruePositive performance at level 5 reflects low frequency (0–3.90Hz)
Sensitivity 100 (15)
= TruePositive FalseNegative × and above low frequency (3.90–7.81Hz) cognitive domination
+
TrueNegative tasks.MLPandNBclassifieraccuracieswere97.14and89.63%,
Specificity 100 (16)
= TrueNegative FalsePositive × respectively,forapproximatedcoefficientsand91.60and81.07%
+ fordetailedatlevel5.Resultsforotherperformanceparameters,
TruePositve
Precision 100 (17) i.e., sensitivity, specificity, precision, and the Kappa statistic,
= TruePositive FalsePositive ×
+ were also impressive. Subject-wise classification accuracies for
P PC
Kappa(k) (cid:0) o− e(cid:1) (18) complex cognitive task (RAPM) vs. eyes open (EO) baseline
= (1 PC)
− e are presented in Tables4, 5. The mean and standard deviation
of accuracies with KNN, SVM, MLP, and Naïve classifiers for
Where P represents the probability of overall agreement
o
between label assignments, classifier and true process; and PCe app1r.6o0x,im97a.t1i4on c1o.e4ffi3,caienndts96(0.0–73.901H.1z9);aanred9d6e.t2a5ile±dc2o.4e7ffi,c9i7en.1t4s
denotes the chance agreement for all labels; i.e., the sum of the ± ± ±
(3.90–7.81Hz)are94.10 2.69,95.35 2.62,97.32 1.42,and
proportionofinstancesassignedtoclassmultipliesinproportion ± ± ±
93.92 2.83,respectively.
totruelabelsofthatspecificclassinthedataset. ±
Classification Results ResultsofDatasetII
WeusedDWTtoextractrelativewaveletenergyfeaturesforD1– For classification of two cognitive tasks, i.e., mental
D5 and A5 for both EEG datasets, where dataset I contained multiplication task and mental letter composing task, MLP
two conditions (EO and RAPM) and dataset II contained two andSVMclassifiersachieved89.17%and86.67%accuracieseach
cognitivetasks,i.e.,mentalmultiplicationtaskandmentalletter forA5approximationcoefficients,andKNNandMLPclassifiers
FrontiersinComputationalNeuroscience|www.frontiersin.org 8 November2017|Volume11|Article103
Aminetal. EEGSignalsAnalysis
FIGURE8|(A)Distributionswithpartialoverlapbetweenpdfsofbothclasses—(eyesopenandRAPM);(B)CorrespondingROCCurvewhere0denotescomplete
overlapand1indicatescompleteseparation.
TABLE2|Approximationcoefficients(0–3.90Hz)forcognitivetasks. TABLE4|Approximationcoefficients(0–3.90Hz)forsubjectwiseclassification
accuracy.
Classifier Accuracy Sensitivity SpecificityAUC Precision Kappa
(%) (%) (%) (%) Statistic KNN SVM MLP BN
KNN,k 3 98.21 96.88 99.63 0.98 99.64 0.96 Subject1 97.14 100 97.14 97.14
=
SVM,RBF 99.11 99.28 98.93 0.99 98.93 0.97 Subject2 98.57 98.57 98.57 95.71
MLP,N 5 97.14 96.48 97.83 0.98 97.86 0.94 Subject3 94.28 97.14 94.28 95.71
=
Naïve 89.63 88.24 90.77 0.94 91.07 0.78 Subject4 92.85 95.71 95.71 94.28
Subject5 92.85 97.14 98.57 95.71
Relativewaveletenergyclassificationresultsforlevel5.
Subject6 100 98.57 97.14 95.71
Subject7 97.14 95.71 97.14 95.71
TABLE3|Detailedcoefficients(3.90–7.81Hz)forcognitivetasks. Subject8 97.14 100 98.57 98.57
Classifier Accuracy Sensitivity SpecificityAUC Precision Kappa Mean 96.25 97.86 97.14 96.07
(%) (%) (%) (%) Statistic
STD 2.47 1.60 1.43 1.19
KNN,k=3 98.39 98.22 98.56 0.98 98.56 0.96 Relativewaveletenergyclassificationresultsforlevel5.
SVM,RBF 98.57 97.89 99.28 0.98 99.28 0.96
MLP,N 5 91.60 88.70 94.98 0.92 95.36 0.89
= activecognitivestatesofconsciousness.UsingSVM,MLP,KNN,
Naïve 81.07 82.71 79.59 0.82 78.57 0.79
and NB classifiers for both conditions, we classified extracted
Relativewaveletenergyclassificationresultsforlevel5. relativewaveletenergyfeaturesforD2–D5andA5.Classification
resultswerenotprominentlyevidentforalldecompositionlevels.
Results for the relative energy of approximation and detailed
achieved93.33and92.50%accuraciesforD5detailscoefficients,
coefficients at level 5 showed the highest performances for
respectively,asshowninTables6,7.Theseresultsindicatethat
cognitive task dominations at low frequency (0–3.90Hz) and
the proposed approach is strong enough to perform on public
abovelowfrequency(3.90–7.81Hz;Tables2,3,7,8).
dataset for classification of cognitive tasks. Comparison of the
These results indicate that relative wavelet energy for low
proposed approach with previous work on the same dataset is
frequency (0–3.90Hz) and above low frequency bands (3.90–
presentedinTable8.
7.81Hz) is a useful feature for EEG classification of cognitive
tasksaswell asseparationof baseline(eyesopen)and complex
DISCUSSION cognitivetask.TheseresultsconfirmDWT’sabilitytocompactly
represent EEG signals and compute total and relative energy
This study used a “pattern recognition” based approach to levels for different frequency bands. The normalization process
classify EEG signals that were recorded during resting and reducedthenon-Gaussianityofextractedfeatures.Furthermore,
FrontiersinComputationalNeuroscience|www.frontiersin.org 9 November2017|Volume11|Article103
Aminetal. EEGSignalsAnalysis
TABLE5|Detailedcoefficients(3.90–7.81Hz)forsubjectwiseclassification (1990)andapplieddifferentfeatureextractionandclassification
accuracy. algorithms. Here, the authors reported the results of previous
studies for comparison of only two cognitive tasks, i.e., mental
KNN SVM MLP BN
multiplicationandmentallettercomposing,forexample81.5%
Subject1 90.00 91.42 97.14 91.42 classification accuracy reported by Keirn and Aunon (1990)
Subject2 95.71 97.14 95.71 98.57 for classification of mental multiplication from mental letter
Subject3 94.28 95.71 95.71 95.71 composing(seeTable8foradetailedsummaryofthesemethods
and results as reported in the literature). The cited authors
Subject4 90.00 98.57 98.57 92.85
reported low classification accuracy rates and used complex
Subject5 94.28 95.71 97.14 91.42
classification models such as, neural network or kernel-based
Subject6 95.71 91.42 100 91.42
classifiers.Ourproposedscheme,asdemonstratedinthepresent
Subject7 95.71 95.71 97.14 92.85
study, yielded higher accuracy rates with a SVM and KNN,
Subject8 97.14 97.14 97.14 97.14
indicatingsuperiordiscriminatoryperformanceswhenassessing
Mean 94.10 95.35 97.32 93.92 mental tasks. In addition, the present study achieved high
STD 2.69 2.62 1.42 2.83 classification accuracy than our previous study using the same
EEGdataset.Therefore,theproposedmethodemployedfeature
Relativewaveletenergyclassificationresultsforlevel5.
normalization and optimization modalities that appear to have
yieldedamoreefficientandreliablesolution.
TABLE6|Approximationcoefficients(0–3.90Hz)forcognitivetasks(mental
multiplicationvs.mentallettercomposing).
LIMITATIONS
Classifier Accuracy Sensitivity SpecificityAUC Precision Kappa
(%) (%) (%) (%) statistic There are some limitations in the present study, which should
be highlighted for future research. The authors used dataset
KNN,k 3 82.50 80.95 84.21 0.85 85 0.84
= from their previous study and a small public dataset. However,
SVM,RBF 86.67 87.93 85.48 0.89 85 0.87 alargepublicdatasetcanvalidatetherobustnessoftheproposed
MLP,N=5 89.17 89.83 88.52 0.90 88.33 0.89 methodforEEGsignalsclassification.Inaddition,EEGrecorded
Naïve 78.33 77.42 79.31 0.79 80 0.77 during a cognitive task is relatively easy to separate from
a baseline EEG than EEG signals recorded in two different
Relativewaveletenergyclassificationresultsforlevel5.
cognitivetasks.Infuturestudy,wewillextendtheapplicationof
theproposedmethodtoclinicaldatasets,suchas,classificationof
TABLE7|Detailedcoefficients(3.90–7.81Hz)forcognitivetasks(mental ictalvs.inter-ictalornormalEEGpatterns.
multiplicationvs.mentallettercomposing).
Classifier Accuracy Sensitivity SpecificityAUC Precision Kappa CONCLUSION
(%) (%) (%) (%) statistic
Theauthorspresentedapatternrecognitionbasedapproachfor
KNN,k 3 93.33 91.94 94.83 0.92 95 0.93
= classification of cognitive tasks using EEG signals. The DWT
SVM,RBF 90 87.50 92.86 0.92 93.33 0.90
was applied to EEG signals for decomposition. Classification
MLP,N 5 92.50 90.48 94.74 0.94 95 0.91
= results were superior to our previous study (Amin et al.,
Naïve 84.17 82.54 85.96 0.85 96.67 0.83
2015a) for dataset I as well as previous work done on dataset
Relativewaveletenergyclassificationresultsforlevel5. II. The experimental results validated the proposed scheme.
These outcomes suggest promising potential for the method’s
application to clinical datasets as a beneficial adjunct for
the use of the proposed feature selection approach minimized
discriminationbetweennormalandabnormalEEGpatterns,as
non-significantfeaturesfroma“features”setbeforefeedingitto
it is able to cope with variations in non-stationary EEG signals
classifiers,whichreducescomputationalcost.
viathelocalizationcharacteristicoftheWT(Rossoetal.,2001).
ResultsofDatasetIarenotexactlycomparablewithprevious
Finally, in this study, the combination of DWT with FDA and
work except the authors’ own work published in 2015 in
PCAtechniquesprovidearobustfeatureextractionapproachfor
which the same EEG data was used as described in dataset
classificationofcognitivetasksusingEEGsignals.
I. However, dataset II is a publicly available dataset, which
is used to compare with previous work done on the same
dataset.Thenotionreflectedbythestudiesisthediscriminatory ETHICS STATEMENT
performance levels (accuracy) achieved by various quantitative
analyticalapproachestoclassifytwodifferentcognitiveactivities ThisresearchstudywasapprovedbytheResearchCoordination
using EEG signals, for example mental multiplication and Committee of Universiti Teknologi PETRONAS, Malaysia. All
mental letter composing(Liang et al., 2006; Zhang et al., the participants had signed the informed consent form before
2010; Vidaurre et al., 2013; Dutta et al., 2018). These studies starting the experiment. The consent document had a brief
employedtheEEGdataoriginallyreportedbyKeirnandAunon descriptionofthisresearchstudyconcerninghumans.
FrontiersinComputationalNeuroscience|www.frontiersin.org 10 November2017|Volume11|Article103