Harmonic analysis of steady-state visual evoked potentials in brain computer interfaces

Harmonic

analysis

of

steady-state

visual

evoked

potentials

in

brain

computer

interfaces

Volkan

C¸

etin

_,

_Serhat

_Ozekes

_,

_Hüseyin

_Selc¸

_uk

_Varol

a_{DepartmentofComputerEngineering,ArelUniversity,34722Tepekent,Istanbul,Turkey} b_{DepartmentofComputerEngineering,UskudarUniversity,34662,Uskudar,Istanbul,Turkey}

c_{DepartmentofElectronicsandCommunicationEngineering,DogusUniversity,34722Kadıkoy,Istanbul,Turkey}

a

r

t

i

c

l

e

i

n

f

o

Articlehistory:

Received22September2019

Receivedinrevisedform28February2020 Accepted30April2020

Availableonline15May2020 Keywords:

Brain-computerinterface

Steady-statevisualevokedpotentials Harmonicanalysis

Featureextraction

a

b

s

t

r

a

c

t

Thesignalsgeneratedintheoccipitallobeofthebrain,asaresultofvisualstimuliflickeringatacertain frequency,arecalledsteady-statevisualevokedpotentials(SSVEPs).SpectralpropertiesoftheSSVEPsare extractedtouseinclassificationstageinSSVEPbasedbrain-computerinterfaces.However,therehasbeen nopreviousstudyexaminingtheeffectsofSSVEPharmonicsonclassificationperformance.Thefrequency spectrumofanSSVEPconsistsofharmonicsatfrequenciesthatareintegermultiplesofthestimulussignal frequency.Inthisstudy,theeffectsofthefirstfourharmonicsofSSVEPsinclassificationperformance wereinvestigated.Totalandrelativebandpowervalues,extractedfromvariouscombinationsofthefirst fourharmonicsofSSVEPs,areusedasfeatures.Duetothequasi-sinusoidalnatureofSSVEPs,ithasbeen observedthattheclassificationmadebythefeaturesextractedfromthesecondharmonicgivesbetter resultsthantheclassificationmadebythefeaturesextractedfromthefirstharmonic.Inaddition,ifmore thanoneharmonicisusedinfeatureextraction,itwasobservedthatthebestclassificationperformance wasobtainedwiththepropertiesextractedfromthesetof1st,2ndand4thharmonics,inalmostall cases.Furthermoreastatisticalstudywasperformedbyapplyingvarianceanalysistotheobtaineddata toverifythesignificanceoftheresults.

©2020ElsevierLtd.Allrightsreserved.

1. Introduction

Brain-ComputerInterface(BCI)isdefinedastheinterfacethat usesbrainsignalstocontroladevice,ortoprovidecommunication betweenthedeviceandtheuser[1].Amorecomprehensive def-initionforBCIisthatthemediuminwhichtheelectricalactivity generatedbythebrainistransmittedtothenervesandmuscles arounditindependentlyofnormalexitpathways[2].BCIdesign canbenefitfromoneormoreelectrophysiologicalsourcesrecorded fromvariousregionsofthebrain.Asaresultofavisualstimulus, theelectricalsignalsseenintheoccipitalandparietallobesofthe brainarecalledvisualevokedpotentials.TheVEPobtainedfrom visualcortexinconsequenceofstimuliatfrequenciesbelow3.5 HziscalledtransientVEP[3,4],becausethestimuluscannot trig-gertogenerateacontinuoussinusoidal-likeresponseinthevisual cortex.Atstimulusfrequenciesbetween3.5Hzand75Hz,a quasi-sinusoidalwaveformisformedduetosuperimpositionoftheaction

∗ Correspondingauthor.

E-mailaddresses:volkancetin@arel.edu.tr(V.C¸etin),

serhat.ozekes@uskudar.edu.tr(S.Ozekes),hsvarol@dogus.edu.tr(H.S.Varol).

potentialsgeneratedinthevisualcortex[5].Duringthepresenceof thestimulus,thevisualcortexgeneratesaquasi-sinusoidalsignal. Therefore,theVEPobtainedinconsequenceofstimuliatfrequency of3.5Hzandaboveisdefinedasthesteady-stateVEP(SSVEP)[4]. InaSSVEP-basedBCIdesign,stimulimustbebetween3.5Hzand 75 Hzsothata classificationcanbepossibleusingthespectral propertiesofthesignals.

Lalor et al. showed that control of computer games can be accomplishedbybinaryclassificationofEEGreceivedfromO1and O2electrodesinaSSVEP-basedBCIsystem[6].Theyusedtwo stim-uliatfrequencies6Hzand25Hz,andextractedfeaturesfromthe powerspectraldensity(PSD)ofSSVEPforclassification.Kellyetal. madeabinaryclassificationofSSVEPsbyusingtwostimuliat fre-quencies10Hzand 12Hz [7]. TheyrecordedEEGfromO1and O2electrodes,andmadeaclassificationbasedonPSDofSSVEPs. Muller-PutzandPfurtschellershowedthatafour-taskclassification byaSSVEP-basedBCIispossibletocontrolabiaxialhand prosthe-sis[8].PruecklandGugerperformedafour-taskclassificationof SSVEPsusingfourstimuliatfrequencies10Hz,11Hz,12Hzand 13Hz[3].TheyusedfirstandsecondharmonicsofPSDofSSVEPs recordedfromO1,O2,Oz,PO3,PO4,PO7,PO8andPOzelectrodes. Binetal.madeacanonicalcorrelationanalysis(CCA)based classifi-https://doi.org/10.1016/j.bspc.2020.101999

cationinaSSVEP-basedBCI[9].TheyusedEEG,recordedfromnine electrodesintheoccipitalandtemporalregions.LuoandSullivan madeafour-taskclassificationbyaSSVEP-basedBCIusingonlythe PO2electrode[5].Theyusedfourstimuliatfrequencies9Hz,10 Hz,11Hzand12Hz,andmadeaclassificationbasedonPSDof SSVEPs.VolosyakI.developedanSSVEP-basedBCI,usingfive stim-uliatfrequenciesof6.67Hz,7.5Hz,8.57Hz,10Hzand12Hz,and madeaclassificationbasedonPSDofSSVEPs[10].Longetal. con-trolledabattery-poweredchairusingahybridBCIapproach.While usingsensorimotorEEGfordirectionanddecelerationofthechair, theyusedSSVEPsobtainedfromO1,O2andOzelectrodesfor accel-eration[11].Leeetal.controlledamobilerobotusingstimuliat frequencies13Hz,14Hzand15Hzthatarecorrespondingtothree differenttasks[12].Zhangetal.designedanSSVEP-basedBCIwith aCCA-basedclassification.Theyusedstimuliatfrequencies6Hz, 7Hz,8Hzand9Hz[13].TheydesignedtheirBCIusingthesignals recordedfromtheO1,O2,Oz,P7,P8,P3,P4andPzelectrodes.Chen etal.controlledaroboticarmbySSVEPs,usingthesignalsrecorded fromPz,PO5,PO3,POS,PO4,PO6,O1,Oz,O2electrodes[14].Park etal. used theextension of multivariatesynchronization index (EMSI)algorithmoverthreeharmonicsofthefrequencyspectrum ofSSVEPsintheirstudywhere theyusedstimuliatfrequencies 7.5Hz,8.57Hz,10Hzand12Hz[15].Choietal.investigatedthe classificationperformanceandeaseofuseofBCIsystemsina vir-tualrealityenvironment[16].TheyrecordedEEGfromCz,PO3,POz, PO4,O1,Oz,andO2electrodes,andusedEMSIalgorithmforfeature extraction.TherearemanySSVEP-basedBCIdesignsandstudiesin theliterature.Inthesestudies,PSD,CCAorEMSIbasedmethods wereappliedtothefrequencyspetrumofSSVEPs.However,there isnostudyinliterature,investigatinghowtheharmonicsofSSVEPs affecttheclassificationperformance.Inthisstudy,the classifica-tionperformedwiththefeatures extractedfromtheharmonics obtainedfromthefrequencyspectrumofSSVEPswerecompared. Thefeaturevectorsconsistofthetotalandrelativebandpower val-uesofthevariouscombinationsofthefirstfourharmonicsobtained fromthePSDoftheSSVEPs.Itisobservedthattheamplitudeof thesecondharmonicishigherthanthefirstharmonicinalmost allcases.Theamplitudesofthethirdandfourthharmonicswere observedtobelowasexpected,anddoesnothaveaneffectthat canimprovetheclassificationperformanceatanacceptablerate.

2. Materialsandmethods

EmotivEpocwasusedtorecordtheEEGinthisstudy.There are 16 electrodes onthe head-set,two of which are reference electrodes.Theactiveelectrodesintheneuro-headsetareinfixed positionsasAF3,F7,F3,FC5,T7,P7,O1,O2,P8,T8,FC6,F4,F8,AF4. P3andP4areusedasreferenceelectrodes.SSVEPsthatarerecorded fromtheoccipitalregionofthebrainwereusedasthe electroneu-rophysiologicalsourceoftheBCI.TheheadsethasonlyO1andO2 channelsontheoccipitallobes.Therefore,onlydatarecordedfrom thesechannelscouldbeused.PruecklandGugerhavetried30

dif-Table1

FrequencysetofstimuliusedtotriggerSSVEPsinHertz.

Stimuli 1stHarmonic 2ndHarmonic 3rdHarmonic 4thHarmonic

1st 4 8 12 16

2nd 4.6 9.2 13.8 18.4

3rd 5.3 10.6 15.9 21.2

ferentchannelcombinationsonfivedifferentsubjectsinorderto determinethechannelsfromwhich thebestSSVEPresponse is obtained.TheyappliedsurfaceLaplaciantransformationsasa mea-surementstrategy.TheirresultsshowthatO1,O2andOzarethe bestpossiblechannelcombinationfromwhichtheSSVEPresponse isobtained[3].ThissuggeststhattheO1andO2channelslocated onEpocaresufficientforaSSVEP-basedBCIapplication.

EEGrecordingswereobtainedfromsevendifferentsubjects.Six ofthemwerecompletelyinexperiencedwithBCIwhileone sub-jecthasahistoryofBCIexperience.Atotalof12minofEEGwere recordedforeachsubject.Aninterfacewiththreebuttonsflickering atdifferentfrequencieswasdesigned,andthesubjectswereasked tolookateachofthethreebuttonsforfourminutes.Thismeans 840datasamplesperclassfor2-stimewindows,and420data samplesperclassfor4-stimewindows.Thestimulusfrequencies, eachcorrespondingtoaseparatepseudo-task,aregiveninTable1. Afour-minuteEEGwasrecordedforeachstimulus,andsubjects wererestedforthreeminutesbetweenrecordings.

2.1. Stimulitypeandfrequency

AcomputerprogramwasdevelopedtotriggeraSSVEPinthe visualcortex.Therearethreebuttonsflickeringatthespecified fre-quenciestotriggeraSSVEPasshowninFig.1.Thedistancesofthe buttonstoeachotherandtotheedgesofthemonitorarealsogiven inFig.1.Thesedistancesaremeasuredona21.5monitorwitha resolutionof1920×1080pixels,whentheprogramisinfullscreen. Themonitor’sdistancetothesubjectsissettobeapproximately60 cmduringtherecordings.

ThereisanasymmetryinSSVEPsasshowninFig.2.A signifi-cantdifferenceisobservedbetweentheamplitudeoftheresponse obtainedatthetimeofswitchingfromtheoffpositiontotheon positionandtheamplitudeoftheresponseobtainedatthetimeof switchingfromtheonpositiontotheoffposition.Asymmetryin theluminanceofonandoffpositionsofthestimulusalsocreatesan asymmetryintheresponse.Therefore,theresponseincludesnot onlythefundamentalfrequencyofthestimulus,butalsohigher harmonics[4].Evenandoddharmonicsoffundamentalfrequency areobservedintheasymmetricSSVEPresponse.Ontheotherhand, symmetricSSVEPsonlycontainevenharmonics offundamental frequency[4]. Asymmetric SSVEPresponsecan beobtainedby usingoppositecheckerboardpatternsasstimuli.However,a uni-formsurfacewasusedasastimulus,takingintoaccountthatamore

Fig.2. (a)Stimulussignalasasquarewave,and(b)asymmetricresponseinthe visualcortex.

provocativestimulussuchasapatternwouldcausemorefatigue becauseoflong-standingSSVEPrecordings[5].

The three stimuli shown in Fig. 1 are applied with button controlschangingbackground coloratspecific frequencies.The adjustmentofthefrequenciesofthebuttonsisoneofthedifficulties encounteredintheuseofacomputerprogramwithaSSVEP-based BCI.Whendeterminingthefrequencyofastimulus,thetime reso-lutionoftheoperatingsystem,theprogramminglanguageandthe monitormustbeconsidered.

Monitor-basedtime resolutionconstraintscausemajor prob-lemsindesignofstimulusprogramsinBCIapplications.Onlythe partoftheSSVEPresponsebetween3.5Hzand30Hzismeasurable byamonitorwitha60Hzrefreshrate.Whenassessedintermsof bandwidth,itcanbeconcludedthatthemonitorisnotasignificant restrictionofadjustingthefrequencyofastimuli.However,dueto thetimeresolutionofthemonitor,onlystimuliatcertain frequen-ciescanbeobtainedinasmoothsquarewaveform.Forexample, frequencyofastimuluscannotbeexactly9Hzonamonitorwith arefreshrateof60Hz.Inorderforafrequencyofastimulusto beseenatapreciselysetvalue,thefrequencymustbedivisibleby 60.Inotherwords,thestimulusfrequenciesthatcanbeseenina regularsquarewaveformonamonitorwitharefreshfrequencyof 60Hzare1Hz,2Hz,3Hz,4Hz,5Hz,6Hz,6.67Hz,7.5Hz,8.57Hz, 15Hz,20Hz,and30Hz[17].Ontheotherhand,asaresultofthe spectralanalyzescarriedout,itisobservedthatstimulithatarenot inperfectsquarewaveform,suchas9Hzand11Hz,triggerSSVEPs whosefrequenciesareapproximately9Hzand11Hz.Forinstance, thefirstandsecondharmonicsofa20-slongSSVEPrecord, trig-geredbyastimuluswithafrequencyof4.6Hz,areobservedat4.56 Hzand9.16Hz,respectively,asseeninthefrequencyspectrum showninFig.3.

2.2. Signalpre-processinganddigitalfiltering

Theonlysignalpre-processinginthisstudyistoclearthe direct-current(DC)offsetvaluesthatvarybetween4−5mV.Atfirstglance itcanbeconsideredthattheDCoffsetvaluecanbeclearedwith a high-passfilter (HPF).It is necessary toattenuateby 100dB orhighertocleanthehighamplitude-lowfrequencynoises[18]. However,DCoffsetcannotberemovedbyapplyingaHPFwitha stopbandattenuationof100dBduetocertainlimitations encoun-teredinthedigitalfilterdesign.Averystrongattenuation,suchas

Fig.3. FrequencyspectrumofSSVEPthatcontainsfirstandsecondharmonicsat thefrequenciesof4.56Hzand9.16Hz.

Table2

CharacteristicsofHPFandLPFappliedinfilteringEEGsignals.

Filtercharacteristic HPF LPF Attenuation 37dB 60.5dB PassbandRipple 0.33dB 0.1dB TransitionWidth 4.6Hz 12Hz 6dBPoint 2.48Hz 38Hz 3dBPoint 3.22Hz 36.7Hz FilterLength 45 42

100dB,willcauseatimedelaywellabovetheacceptablevalues. Forthisreason,theDCoffsetisremovedbytakingthedifference betweenthesignalanditsarithmeticmean.Inelectrophysiology ButterworthIIR filters or FIRfilters areusually applied[18]. In thisstudy,FIRfilterdesign wasappliedinordertoavoid signal distortionsespeciallyduetononlinearphase.Table2showsthe characteristicsoftheHPFandlow-pass(LPF)filters.

2.3. Featureextractionandclassification

ThefeaturesusedinclassificationofEEGaregenerally classi-fiedintotimedomainandfrequencydomain.Timedomainfeatures usedinBCIstudiescanbewaveform-based,suchaspeakvalueof thesignalorblock-basedsuchasaverageoveraspecifiedtime win-dow[19].TherearealsoBCIstudiesinvolvingtemplatematching asfeatureextraction.Inthesestudies,thesimilarityoftheEEGwith apredefinedtemplatewasusedasafeature[19].BCIusersarebest atmodulatingthespectralcharacteristicsofthestimuli.Hence,the spectralfeaturesofEEGsignalsaremostlyusedinSSVEP-based BCIapplications[3,6,10,12,19–21].Inthisstudythefeaturevectors consistofthetotalandrelativebandpowervaluesofthevarious combinationsofharmonicsobtainedfromthefrequencyspectraof theSSVEPs.

Itisobservedthatthemostwidelyusedclassificationmethods inBCIdesignareartificialneuralnetworks(ANN),supportvector machines(SVM)andNaiveBayes(NB)[2,22].Inthisstudy,binary andternaryclassificationswereperformed.Three-layered percep-tronnetwork,trainedbyback-propagationalgorithm,wasusedin theclassificationofEEGsignals.Theinputandtheoutputlayers containneuronsasmanyasthenumberoffeaturesandnumberof classes,respectively.Numberofneuronsinhiddenlayerissetto bethemeanofthenumberofneuronsintheinputandtheoutput layers.Thelearningrateofthenetworkis0.3andthenumberof trainingiterationsis40.Gridsearchandcrossvalidationwereused forhyper-parameteroptimizationofANNandDVMalgorithms.A 10-foldcross-validationmodelwasusedtoevaluatethe classifica-tionresults.Eachfoldiscomposedofrandomsampleswithequal classdistribution.

Fig.4. BinaryClassificationAccuracyforseveralharmonicsetsfor2sTime-Windows.

3. Resultsanddiscussion

Thisstudyhasshownthatthemostprominentfeaturescanbe extractedfromthesecondharmonicoftheSSVEPresponse.Ithas alsobeenobservedthatthefirstandsecondharmonicstogether aremoreprominentthananyotherbinarysetofharmonics com-binations.Inaddition,1st,2ndand4thharmonic-basedfeatures providebetterclassificationperformancethan1st,2ndand3th harmonic-basedfeatures,albeitwithaslightdifference.The accu-racypercentages of binaryand ternaryclassifications using2-s windowlengthforallsubjectsaregiveninFigs.4and5, respec-tively.Whenboththebinaryandternaryclassificationresultsgiven inFigs.5and6areexamined,itisobservedthattheeffectofthefirst harmonicontheclassificationperformancealoneisinadequate.On theotherhand,intheclassificationswheresecond,thirdand/or fourthharmonicsareusedtogetherwiththefirstharmonic,the accuracyrateincreasesdrastically.

Wealsoperformedbinaryandternaryclassificationswith4-s windowlength.Indoingso,wewantedtoseetheeffects ofthe featuresextractedfromdifferentharmonicsetsonthe classifica-tionperformancemoresignificantlybycollectingpropertiesover alongerperiodoftime.Theaccuracypercentagesofbinaryand ternaryclassificationsusing4-swindowlengthforallsubjectsare giveninFigs.6and7,respectively.Intheclassificationsmadewith 4-swindowlength,parallelresultswereobtainedwiththosein2-s windowlength.

Whenexaminingtheaverageperformancesofclassifications basedononlyfirstandonlysecondharmonics,theresultsshow that the second harmonic-based features are more prominent than the first harmonic-based features. Therefore, the second harmonic-basedfeaturesshouldbepreferredifonlysingle har-monicissufficientforclassification.Asexpected,theclassifications basedonthefirsttwo harmonicstogetherisfarsuperiortothe classificationsbasedonthefirstandsecondharmonicsalone, inde-pendentofthesubject,varietyoftasks,andthewindowlength. It is also understood that in all cases in which the first and secondharmonicsareusedtogether,theclassificationsperform betterthanthecasesinwhichthefirsttwoharmonicsareused alone.

Itis clearthat inthe caseof usingthree harmonic frequen-ciestogether,thefirstandsecondharmonicsaremoreprominent. Normally,thirdharmonicisconsideredtobemorepowerfulthan thefourthharmonic.However,asaresultofthethisstudy,itwas observedthatthefourthharmonic-basedfeaturesaremore promi-nentthanthethirdharmonic-basedfeatures.Thisresultisbelieved

Table3

SummaryoftheSSVEPHarmonicAnalysis.

Harmonics Meanaccuracy

2-s 4-s

2-Class 3-Class 2-Class 3-Class

1 73,6 53,6 78,1 56,0 1,2 85,5 68,9 88,7 74,5 1,2,3 86,1 71,6 88,9 76,1 1,2,4 86,5 71,6 89,8 75,4 1,2,3,4 85,6 70,6 88,1 74,4 2 82,4 66,3 87,3 71,8

tocausedbecauseoftheasymmetricresponseofthevisualcortex totheflickeringstimulus.

Table3showsthemeanclassificationaccuraciesofallharmonic sets.InTable3,theresultsobtainedfromharmonicsetssuchas 2,3,4or3,4arenotincluded.Thereasonforthisisthatthe classi-ficationaccuracywiththefeaturesextractedfromtheseharmonic setsisverylow.Bestclassificationperformancesarehighlightedas boldcharacters.Itisseenthat1st,2ndand4thharmonics-based featuresarethebestpreferencesinclassificationofSSVEPsinterms ofclassificationaccuracy.

Analysisofvariance(ANOVA)isappliedtoverifywhetherthere isastatisticalsignificanceinclassificationaccuraciesofdifferent harmonic sets. When the accuracy percentages of binary clas-sification using 2-s window length were analyzed by one-way ANOVA,astatisticalsignificance betweenharmonicgroups was observed(F(5,36)=3.164,p=.018).Astatisticalsignificancewas alsoobservedbetweenharmonicsetsofternaryclassificationusing 2-swindowlength(F(5,36)=4.067,p=.005).Itisobservedthata statisticalsignificancebetweenharmonicsetsexistsafteranalyzing theaccuracypercentagesofbinaryclassificationusing4-swindow length,(F(5,36)=3.355,p=.014).Finally,asignificantdifference wasobservedbetweenharmonicsetsfortheaccuracypercentages ofternaryclassificationusing4-swindowlength(F(5,36)=7.117, p=.000).Theseone-wayANOVAresultsshowthestatistically sig-nificantdifferencesinmeanclassificationaccuracyvaluesbetween differentharmonicsets.

Tukeytestwasusedforconductingposthoctestsonone-way ANOVAtodeterminestatisticallysignificantdifferencesbetween harmonicsets.Tukeyposthoctestrevealedthatforbinary clas-sificationusing2-swindowlength,theaccuracyoffirstharmonic (73.6±10.5)wasstatisticallysignificantlylowerthanfirst-second harmonics(85.5±6.9,p=.049),first-second-thirdharmonics(86.1

Fig.5.TernaryClassificationAccuracyforseveralharmonicsetsfor2sTime-Windows.

Fig.6. BinaryClassificationAccuracyforseveralharmonicsetsfor4sTime-Windows.

Fig.7.TernaryClassificationAccuracyofSubjectsfor4sTime-Windows.

±6.0,p=.033),first-second-fourthharmonics(86.5±5.9,p=.027), andfirst-second-third-fourthharmonics(85.6±6.0,p=.047).For ternaryclassificationusing2-swindowlength,theaccuracyoffirst harmonic(53.6±12.5)wasstatisticallysignificantlylowerthan first-secondharmonics(68.9_±8.3,p=.34),first-second-third

har-monics(71.6±8.4,p=.008),first-second-fourthharmonics(71.6 ±8.3,p=.009),andfirst-second-third-fourthharmonics(70.6± 7.7,p=.014).Forbinaryclassificationusing4-swindowlength,the accuracyoffirstharmonic(78.1±9.3)wasstatisticallysignificantly lowerthanfirst-secondharmonics(88.7_±5.9,p=.33),

first-second-Inpreviousyears,firstandsecondharmonicsbasedfeaturesare usedforfeatureextractioninSSVEP-basedBCIstudies.However, therehasbeennopreviousstudyexaminingtheeffectsofSSVEP harmonicsonclassificationperformanceinBCIs.Tothebestofour knowledge,thisisthefirststudythatinvestigatedtheeffectsofthe frequencyharmonicsofSSVEPsinclassificationperformance. In thisstudy,theclassificationperformedwiththefeaturesextracted fromthedifferentsetsofharmonicsofSSVEPswereinvestigated. Thefeaturevectorsconsistofthetotal andrelativebandpower valuesoftheharmonics.Itisobservedthatclassification perfor-manceis betterwithsecond harmonic-based features thanthe classificationperformanceofthefirstharmonic-basedfeaturesin almostallcases.Theamplitudesofthethirdandfourth harmon-icswereobservedtobelowasexpected.However,ifthefourth harmonic-basedfeaturesareusedtogetherwiththe1stand2nd harmonics,itisseenthattheclassificationperformanceincreases slightly.

Authors’contribution

VolkanC¸etin: Data curation, Formal analysis, Investigation,

Software,Methodology,Writingoriginaldraft,Writingreview.

Ser-hatOzekes:Conceptualization,Investigation,Validation.Hüseyin

Selc¸ukVarol:Conceptualiztion,Supervision,Project

administra-tion.

DeclarationofCompetinginterest

Theauthorsdeclarethattheyhavenoknowncompeting finan-cialinterestsorpersonalrelationshipsthatcouldhaveappearedto influencetheworkreportedinthispaper.

References

[1]J.VanErp,F.Lotte,M.Tangermann,Brain-computerinterfaces:beyond medicalapplications,Computer45(4)(2012)26–34.

attentioncontrolinanindependentbrain-computerinterface,IEEETrans. Biomed.Eng.52(9)(2005)1588–1596.

[8]G.R.Muller-Putz,G.Pfurtscheller,Controlofanelectricalprosthesiswithan SSVEP-basedBCI,IEEETrans.Biomed.Eng.55(1)(2008)361–364.

[9]G.Bin,X.Gao,Z.Yan,B.Hong,S.Gao,Anonlinemulti-channelSSVEP-based brain-computerinterfaceusingacanonicalcorrelationanalysismethod,J. NeuralEng.6(4)(2009),046002.

[10]I.Volosyak,SSVEP-basedBremen-BCIinterface-boostinginformation transferrates,J.NeuralEng.8(3)(2011),036020.

[11]J.Long,Y.Li,H.Wang,T.Yu,J.Pan,F.Li,Ahybridbraincomputerinterfaceto controlthedirectionandspeedofasimulatedorrealwheelchair,IEEETrans. NeuralSyst.Rehabil.Eng.20(5)(2012)720–729.

[12]P.Lee,H.Chang,T.Hsieh,H.Deng,C.Sun,Abrainwaveactuatedsmallrobot carusingensembleempiricalmodedecomposition-basedapproach,IEEE Trans.Syst.ManCybern.PartASyst.Hum.42(5)(2012)1053–1064.

[13]Y.Zhang,G.Zhou,J.Jin,X.Wang,A.Cichocki,SSVEPrecognitionusing commonfeatureanalysisinbrain-computerinterface,J.Neurosci.Methods 244(2015)8–15.

[14]X.Chen,B.Zhao,Y.Wang,X.Gao,Combinationofhigh-frequency

SSVEP-basedBCIandcomputervisionforcontrollingaroboticarm,J.Neural Eng.16(2019),026012.

[15]S.Park,H.Cha,C.Im,Developmentofanonlinehomeappliancecontrol systemusingaugmentedrealityandanSSVEP-basedbrain–computer interface,IEEEAccess7(2019)163604–163614.

[16]K.Choi,S.Park,C.Im,Comparisonofvisualstimuliforsteady-statevisual evokedpotential-basedbrain-computerinterfacesinvirtualreality environmentintermsofclassificationaccuracyandvisualcomfort,Comput. Intell.Neurosci.2019(2019),9680697.

[17]F.T.Hvaring,A.H.Ulltveit-Moe,Acomparisonofvisualevokedpotential (VEP)-basedmethodsforthelow-costemotiveEPOCneuroheadset,in:Msc. Thesis,NorwegianUniversityofScienceandTechnology,Departmentof ComputerandInformationScience,Trondheim,Norway,2014.

[18]A.Widmann,E.Schröger,B.Maess,Digitalfilterdesignfor

electrophysiologicaldata-apracticalapproach,J.Neurosci.Methods250 (2014)34–46.

[19]J.R.Wolpaw,E.W.Wolpaw,Brain-ComputerInterfacesPrinciplesand Practice,OxfordUniversityPress,Inc.,NewYork,USA,2012.

[20]Y.Wang,R.Wang,X.Gao,B.Hong,S.Gao,ApracticalVEP-based

brain-computerinterface,IEEETrans.NeuralSyst.Rehabil.Eng.14(2)(2006) 234–239.

[21]H.Gollee,I.Volosyak,A.J.McLachlan,K.J.Hunt,A.Graser,AnSSVEP-Based brain-computerinterfaceforthecontroloffunctionalelectricalstimulation, IEEETrans.Biomed.Eng.57(8)(2010)1847–1855.

[22]A.Nijholt,D.Tan,Brain-computerinterfacingforintelligentsystems,IEEE Intell.Syst.23(3)(2008)72–79.