• No results found

Sparse time artifact removal

N/A
N/A
Protected

Academic year: 2021

Share "Sparse time artifact removal"

Copied!
7
0
0

Loading.... (view fulltext now)

Full text

(1)

ContentslistsavailableatScienceDirect

Journal

of

Neuroscience

Methods

j ou rn a l h o m epa g e :w w w . e l s e v i e r . c o m / l o c a t e / j n e u m e t h

Basic

Neuroscience

Sparse

time

artifact

removal

Alain

de

Cheveigné

a,b,c,∗

aLaboratoiredesSystèmesPerceptifs,UMR8248,CNRS,France bDépartementd’EtudesCognitives,EcoleNormaleSupérieure,France cUCLEarInstitute,UnitedKingdom

h

i

g

h

l

i

g

h

t

s

•TheSTARalgorithmaddresseschannel-specificnoisethatissparseintime.

•Itremoveselectrodeorsensornoiseandcertainformsofmyogenicartifact.

•Incontrasttoothertechniques,fewdataarelostandthedimensionalityofthedataispreserved.

•TheSTARalgorithmcomplementscomponentanalysistechniquessuchasICA.

a

r

t

i

c

l

e

i

n

f

o

Articlehistory:

Received25October2015 Receivedinrevisedform 23November2015 Accepted2January2016 Availableonline8January2016

Keywords: EEG MEG LFP ECoG Artifact Myogenic ICA Sensornoise

a

b

s

t

r

a

c

t

Background:MuscleartifactsandelectrodenoiseareanobstacletointerpretationofEEGandother elec-trophysiologicalsignals.Theyareoftenchannel-specificanddonotfullybenefitfromcomponentanalysis techniquessuchasICA,andtheirpresencereducesthedimensionalityneededbythosetechniques.Their high-frequencycontentmaymaskormasqueradeasgammabandcorticalactivity.

Newmethod:Thesparsetimeartifactremoval(STAR)algorithmremovesartifactsthataresparseinspace andtime.Thetimeaxisispartitionedintoanartifact-freeandanartifact-contaminatedpart,andthe correlationstructureofthedataisestimatedfromthecovariancematrixoftheartifact-freepart.Artifacts arethencorrectedbyprojectionofeachchannelontothesubspacespannedbytheotherchannels.

Results:Themethodisevaluatedwithbothsimulatedandrealdata,andfoundtobehighlyeffectivein removingorattenuatingtypicalchannel-specificartifacts.

Comparisonwithexistingmethods:Incontrasttothewidespreadpracticeoftrialremovalorchannel removalorinterpolation,veryfewdataarelost.IncontrasttoICAorotherlineartechniques,processing islocalintimeandaffectsonlytheartifactpart,somostofthedataareidenticaltotheunprocesseddata andthefulldimensionalityofthedataispreserved.

Conclusions:STARcomplementsotherlinearcomponentanalysistechniques,andcanenhancetheir abilitytodiscoverweaksourcesofinterestbyincreasingthenumberofeffectivenoise-freechannels.

©2016TheAuthor.PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBY-NC-ND license(http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction

Amongthemanysourcesofnoiseandartifactthatplague stud-iesinvolvinghumanand animal electrophysiology, someaffect onlyonechannel atatime. Thispaperaddressessuch channel-specificartifacts,leavingasideothertypesthatimpingeonseveral channelssuchaseyeblink,heartbeatorbackgroundneural activ-ity.Signalsrecordedbymultichannelrecordingtechniquessuchas electroencephalography(EEG),magnetoencephalography(MEG),

∗ Correspondenceto:Audition,DEC,ENS,29rued’Ulm,75230Paris,France. Tel.:+33144322672/+447912504027.

E-mailaddress:[email protected]

electrocorticography(ECoG),invasiveelectrodearraysoroptical techniquesarerelatedtounderlyingsources byalinearmixing process:

xj(t)=

i

si(t)uij, (1)

where t is time,si(t)are thesourcesignals, and uij are mixing coefficients.Wecall“channel-specific”a sourcesiforwhichthe uijarezeroforallchannelsjexceptone.

Channel-specificnoiseincludeselectrodecontactnoise, pulsa-tionnoise,andcertainformsofmuscleartifactinEEG,aswellas sensor noise inMEG or othertechniques.Electrode-tissue con-tactartifactsareusuallytemporallysparse,occurringasisolated events or bursts of events that affect one channel at a time. http://dx.doi.org/10.1016/j.jneumeth.2016.01.005

0165-0270/© 2016 TheAuthor. Published by ElsevierB.V. This is an open accessarticle under the CCBY-NC-ND license (http://creativecommons.org/licenses/ by-nc-nd/4.0/).

(2)

Muscleartifactsmayalsobechannel-specificiftheyareproduced bymotorunitsproximaltoasingleelectrode(Yilmazetal.,2014), althoughactivityfromdeepermusclesmayaffectmultiple elec-trodes.Suchartifactsoftenhaveaspectrumthatextendstohigh frequency,buttheymayalsoincludelow-frequencycomponents thatoverlapwithlow-frequencycorticalactivity(Goncharovaetal., 2003),sothatspectralfilteringisnotsufficienttoeliminatethem. Electromyogenicnoiseisparticularlytroublesomeasitcanbe con-fusedwithhigh-frequencycorticalactivity(Yuval-Greenbergetal., 2008;Whithametal.,2007;Muthukumaraswamy,2013;Yilmaz etal.,2014),particularlyasmuscleactivitymaycorrelatewith cog-nitivestate(Whithametal.,2008;Yuval-Greenbergetal.,2008; McMenaminetal.,2011).Cephalicskinpotential(Corbyetal.,1974) mayalsocovarywithcognitivestate,andcontactnoisemay corre-latewithbehavior.

Astandardapproachtodealingwithspatio-temporallysparse artifactsistodiscardeithertheoffendingchannel,orthe offend-ingtimeintervalortrial(Junghöferetal.,2000).Thisentailsloss ofdata,particularlyifartifactsaffectmultiplechannelsand/orare widelydistributedin time,and italsocomplicatesanalysisand interpretationstages,thatneedtobemadetoleranttothemissing data.

Another approach is to apply multichannel linear analysis techniquessuchasindependentcomponentanalysis(ICA), beam-forming,or joint diagonalization (JD) (de Cheveigné and Parra, 2014) to isolate noise components. Component signals yk(t) obtainedbythesemethodsarerelatedtoobservationsas:

yk(t)=

j

xj(t)wjk, (2)

wheretistimeandthewjkareweights.TheJobservationchannels

spanaspacethatcontainsallsuchlinearcombinations,andthe dimensionalityofthedataisthenumberofdimensionsofthisspace (Jor less).Components belongtothisspace.Different methods (PCA,ICA,beamforming,etc.)differinhowtheyfindthe appropri-ateweightstoapplytothedata.ICAinparticularhasbeenproposed toremoveartifactsincludingmyogenic(Delormeetal.,2007;Ma etal.,2012;Crespo-Garciaetal.,2008).

Theappealoftheselineartechniquesisthatanoisesourcexi canpotentiallybeperfectlycanceled:acomponentykis insensi-tivetosourceiaslongas

juijwjk=0,whereuijarethemixing coefficientsandwjkthecomponentweights(Eqs.(1)and(2)).With

high-dimensionaldata(lotsofchannels)thereisconsiderable flex-ibilityin satisfyingthis constraint, and thestrength ofanalysis algorithms suchasICA liesin theirability tofind suchsets of weights.However,ifanoisesourceisspecifictoasinglechannelj, itcanonlybecancelledbysettingwjktozeroforeverycomponent

k,effectivelydiscardingthatchannel.Inthissituation,component analysisofferslittleoverthetime-honoredpracticeofdiscarding noisychannels.

Componentanalysisitselfisvulnerabletochannel-specificnoise becauseitreliesonthedimensionalityofthedata(determinedby thenumberofchannels)toresolvethevarioussources.Ifchannels arediscardedduetoartifacts,analysismaybeimpaired,whereasif theyarenotdiscarded(duetolackofknowledgeortheneedto con-serveenoughchannels),theartifactisinjectedintotheextracted componentsviaEq.(2).Thepresenceofartifactsmayalsointerfere withtheabilityofthealgorithmtofindtheoptimalwjk.For

exam-pleanalgorithmsuchasCSP(Kolesetal.,1990),thatsearchesfor componentsthatdifferinpowerbetweentwointervals,maylock ontoanartifactthatispresentinoneintervalbutnottheother. Finally,theartifactsmayinterferewiththeabilitytoestimatethe topographyassociatedwithcorticalactivity,possibly compromis-ingsourcemodeling.Channel-specificnoisethuslimitstheability

oflinearmethodstoimprovethesignaltonoiseratio(SNR)ofweak brainactivity.

Theseconsiderationsleadustofocusonchannel-specificnoise, leavingothertechniquessuchasICA,JD,orbeamformingtodeal withnoise sourcesthat impinge onmultipleelectrodes.Thisis importantscientifically,toobtainamoreaccuratepictureofbrain activity,andalsoforapplicationssuchasbrain-computerinterfaces (BCI),predictionofepilepticseizures,wearablebrain-monitoring devices,andsoon.Themethoddescribedhereiseffective,fully automatic,andamenabletoanonlineimplementationfor applica-tionsthatinvolverealtimemonitoring.

2. Methods

2.1. Signalmodelandassumptions

Eachobservationxj(t)isthesumofsignalsfrommultiplesources iwithin thebrainor theenvironment (Eq.(1)).We make sev-eralrestrictiveassumptions. Eachnoisesourceni(t)affectsonly oneparticularchannel(assumption1).Noiseactivityistemporally sparsesothatartifactsondifferentchannelsdo nottemporally overlap(assumption 2), andfor a significantproportionof time thedataareartifact-free(assumption3).Finally,weassumethat in theabsence of artifactsthedata are linearlydependent such thateachchannelbelongstothesubspacespannedbytheother channels.Inotherwordsfor eachchanneljthereexistajj such thatxj(t)=

j=/jajjxj(t)(assumption4).Thisisplausiblefor neu-rogenicactivityduetosource-to-sensormixing.Ofcourse,many kindsofnoisedonotmeettheseassumptions;thefocushereison thosethatdo.Inrealdata,theseassumptionswillbemetonlyas anapproximation,forexamplebecauseofnon-stationarityofthe brainandnoiseprocessesunderlyingthedata.Thequalityofthe outcomedependsonthequalityoftheapproximation.

2.2. TheSTARalgorithm

Thealgorithmproceedsintwophases.Thefirstphasedetects thepresenceofchannel-specificartifacts,thesecondphasecorrects them.

Phase1.Thecovariancematrixofthedataisestimated,and fromitiscalculatedthematrix Athatprojectseachchannelon thesubspacespannedbytheotherchannels.Theprojection ¯xj(t)

ofchanneljistheweightedsumofthechannelsj =/ jthatbest approximatesxj(t).Intheabsenceofanartifactweshouldhave ¯

xj(t)−xj(t)=0asaresultofthelineardependenceassumption,soa

significantdeviationindicatesthepresenceofanartifact.Valuesof ¯

xj(t)−xj(t)arefitbyazero-meanGaussiandistribution,andvalues

eccentricfromthisdistribution(relativetoapredefinedthreshold )areflaggedasartifactual.Thisisrepeatedforallchannels,andthe unionofeccentrictimesamplesislabeledasartifact-contaminated. Thecovariancematrixisinitiallyestimatedfromtheentiredata, andsubsequentlyreestimatedonthepartofdatalabeledas artifact-free.Afewiterationsofthisprocessleadtoastablepartitionofthe timeaxisbetweenartifact-freeandartifact-contaminatedparts.

Phase 2. The artifact-contaminated part is further divided accordingtowhichchannelismostdegradedateachtimesample. Forthispurpose,asecondeccentricitymeasureiscalculatedfor eachchannelastheratioofinstantaneouspowertopower aver-agedovertheartifact-freeportion.Thechannelwiththehighest scoreatagiventimesample“owns”thatsample,anditsdataare replacedwithitsprojectiononthesubspacespannedbytheother channels,usingprojectioncoefficientscalculatedfromthe artifact-freepart.Datareplacementoccursonlyatthepartcorresponding totheartifact:atothertimesthedataareleftintact,somostofthe dataremainuntouchedbytheprocessing.

(3)

-1 0 1 A.U. target (a) -20 0 20 A.U.

mixed with noise (SNR=10) and artifact (b) time -20 0 20 A.U. cleaned by STAR (c) -20 0 20 A.U.

mixed with noise (SNR=10-8) and artifacts (d) -20 0 20 A.U. cleaned by STAR (e) -2 0 2 4 A.U. cleaned by JD (f) time -1 0 1 A.U. cleaned by STAR + JD (g)

Fig.1. Denoisingsimulateddata.(a)Targetsignal.(b)10-channeldataincludingtarget,Gaussiannoise(SNR=10),andanartifactononechannel.Theblackbarindicates STAR’sestimateofthe“artifact-contaminated”part.(c)DataafterremovaloftheartifactbytheSTARalgorithm.(d)10-channeldataincludingthesametargetandGaussian noiseatamuchlowerSNR(10−8)andwithtransientartifactsonallchannels.TheblackbarsindicateSTAR’sestimateofthe“artifact-contaminated”part.(e)Sameafter

removalofartifactsbytheSTARalgorithm.(f)ResultofapplyingtheJDalgorithmtotherawdatain(d),withtheaimofrecoveringthetarget.(g)Same,buttheSTARalgorithm wasfirstappliedbeforetheJDalgorithm.

2.3. Implementationdetails

AMatlabimplementationofthealgorithmisavailableinthe NoiseToolstoolbox(http://audition.ens.fr/adc/NoiseTools/).Inthe projectionsteps,toavoidissueswithrank-deficientdata,PCAis appliedandPCswithpowerbelowathresholdarediscarded.The algorithm operates ona sample-by-samplebasis, each channel switchingbetweenan artifact-free state(noprojection) and an artifact-contaminatedstate(projection). Tolimitthenumber of switches,theeccentricitymeasureistemporallysmoothedby con-volutionwithashorttriangularwindow.

Thealgorithmthushas4parameters:theeccentricitythreshold (thesamevalueisusedinphases1and2),thesizeofthe smooth-ingkernelinphase2,thenumberofneighborsusedtofixeach channel,andthetruncationthresholdforthePCA.Ofthese,only thefirstiscritical,asitdeterminestheproportionoftime sam-plesthatwillbelabeledasartifact-contaminated.Anexcessively lowthresholdleadstoinsufficient“artifact-free”datatoreliably determineitscorrelationstructure.For convenience,our imple-mentationisdesignedtoautomaticallyincrementthethreshold (byafactor1.1)ifthisoccurs.

Tosavecomputationandreduceoverfitting,channelsmaybe projectedonasubsetofneighbouringchannels,ratherthanonthe fullsetofchannels.Ontheassumptionthatcorticalsourcesof inter-estshouldaffectgroupsofphysicallyclosechannels,thissubset shouldsuffice.‘Neighborhood’canbequantifiedbasedon geomet-ricalproximity,orbymeasuringbetween-channelcorrelationin therawdata.Toimprovenoiserejectionitmaybeusefultoapply STARrepeatedly,sothateachiterationremovesartifactsmissedby thepreviousiteration.

2.4. Relaxingassumption2

Assumption2is oftenviolatedfor myogenicnoise thatmay occurinburstsaffectingmultiplemotorunits(Yilmazetal.,2014). Wecanrelaxittoallowseveralchannels[jC]tobecontaminated, aslongaswemodifyassumption(4)toassumethateachofthe contaminatedchannels[jC]canbeexpressedastheweightedsum

of the remainingnon-contaminated channels. The algorithm is modifiedtoprojecteachof[jC] onthesubspacespannedbythe non-contaminatedchannels.Thismodificationiscomputationally expensive(thenumberofcasestoconsiderisexponentialinthe numberof[jC]),andinourimplementationitisonlyapproximated, seebelow.

3. Results

3.1. Simulateddata

Asinusoidal targetand severalGaussian noise sourceswere

mixed together to simulate multichannel data, and

channel-specificartifactswereadded.Thesinusoidaltargetsignal(Fig.1(a)) wasmixedintothe10-channeldataviaa1×10mixingmatrixwith randomcoefficients,togetherwith8independentGaussiannoise sourcesmixedviaa8×10randommixingmatrix.Coefficientswere chosentoensureSNR=10,andapulse-likeartifactwasaddedto onechannel(Fig.1(b)).ApplyingSTAReffectivelysuppressesthe artifact(Fig.1(c)),apparentlywithoutdistortingthesignal.Thedata remaincontaminatedbytheothernoisesources,althoughatthis SNRthetargetstillemerges.Fig.1(d)showsasimilarmixturewith SNR=10−8 andoneartifactoneachchannel,andFig.1(e)shows theresultofapplyingSTAR.TheblackbarsinFig.1(b,d)indicate STAR’sestimateoftheartifact-contaminatedpart.Theartifactsare removed,butatthisSNRthetargetisinvisible.

AlinearalgorithmsuchasJD(deCheveignéandParra,2014) canbeusedtoextractaweaktargetsuchasFig.1(a)despiteavery unfavorableSNR.Howeverthisrequiresthedimensionalityofthe noisetobesmallerthanthatofthedata,arequirementthatfails inthepresenceofthechannel-specificartifactsinadditiontothe GaussiannoiseasinFig.1(d).Indeed,applyingJDtothe artifact-contaminateddatafailstorecoverthetarget(Fig.1(f)).However, ifSTAR isfirstappliedtoremove theartifacts(Fig.1(e)),JDcan successfullyrecoverthetarget(Fig.1(g)).Thisexampleshowsthat STARcaneffectivelyremovechannel-specificartifacts,thatitcando soforartifactsaffectingmultiplechannels,andthatsuchremoval

(4)

Fig.2.128-channelEEGdata.(a–c)Examplesofsingle-channelsignalsfroma 128-channelEEGrecordingbefore(red)andafter(blue)applyingSTAR.(d)Percentage ofsamplesremovedforeachchannel(sortedbydecreasingpercentage),foreach subject.(e)Proportionofpowerremovedforeachchannel(sortedbydecreasing power),foreachsubject.(Forinterpretationofthereferencestocolorinthislegend, thereaderisreferredtothewebversionofthearticle.)

maybeneededforacomponentanalysistechniquesuchasJDto succeed.

3.2. EEGdata

Afirstexampleinvolvesdatarecordedwitha128-channelEEG system(Biosemi).Thedatasampledat512Hzwerehigh-pass fil-tered with 20Hz cutoff (2nd order Butterworth) to emphasize thegammaregionthat isrelativelysusceptible toartifacts.The STARalgorithmwasappliedwiththresholdparameter=2to sup-presschannel-specificartifacts.Thesizeofthesmoothingwindow appliedtotheeccentricitymeasure was19 samples.Fig.2(a–c) shows three example signals from individual electrodes before (red)andafter(blue)processing.InFig.2(a)thedataare contam-inatedbywhatseemstobemuscleactivity.Thisissuppressedby STAR(thethreeremainingburstsreflectactivitycorrelatedacross sensorsandthereforenotremoved).InFig.2(b)thedataare con-taminatedbyapulsatilenoisewithaperiodofapproximately1s, mostlikelya“pulsationartifact”duetoaveinunderlyingthe elec-trode(NiedermeyerandLopezdaSilva,2005).InFig.2(c)thedata arecontaminatedbyaseriesofglitchesofhighamplitude(notethe scale)thatareequallywellsuppressed.

Data from this study were available for 8 subjects, about 1.5–2.5heach. Fig. 2(d) shows the percentage of samples cor-rectedbySTAR foreachchannel(channelssortedbydecreasing percentage)and subject, cumulated over therecording session. Thepercentagevaries acrosssubjectsandchannels,reflecting a differentprevalenceofchannel-specificartifactsbetweensubjects, andforagivensubject,betweenchannels.Fortheworstsubjectthe percentageisbelow5%formostchannels,implyingthat>95%of thedataofthosechannelswereuntouchedbytheprocessing.For goodsubjectsthepercentageofintactsamplesis>98%foralmost allchannels.Thesenumberssuggestthatthenegativeimpactof applyingSTAR,ifany,islikelytobeminimal.Fig.2(e)showsthe percentageofpowerremovedbySTARforeachchannel(channels

Fig.3. 128-channelEEGdata.(a)Onechannelwithmuscleartifactbefore(red)and after(blue)processing.(b,c)Spectrogramsofthesignalbeforeandafterapplying STAR.(d,e)Same,withsuperimposedchirpsignal,illustratingthatprocessingcan revealasignalotherwisemaskedbytheartifact.(Forinterpretationofthereferences tocolorinthislegend,thereaderisreferredtothewebversionofthearticle.)

sorted by decreasing percentage) and each subject (individual lines).Againthepercentagevariesacrosssubjectsandchannels.For theworstsubject,>20%ofpowerwasremovedfor>30channels, and>55%fortheworstchannel.Forothersubjectsthe percent-agesaresmaller,butstillsignificantforasubsetofchannels.To summarize,STARisbothsafeandofpotentiallysignificantbenefit. Fig.3(a)showsonechannelcontaminatedbyaburstofpulsatile activity,likelymuscleartifact,before(red)andafter(blue)applying STAR.ThespectrogramofthesedataisshowninFig.3before(b), andafter(c)processing.Artifactualactivitydominatesthehigher spectralregion,butthisisalleviatedbyprocessing.Fig.2(d)and(e) illustrateshowthisprocessingcouldpotentiallyrevealhigh fre-quencyactivity(hereasynthesizednarrowbandchirp)otherwise maskedbytheartifact.

Asecondexampleusesdatarecordedfromthreeclosely-spaced EEGelectrodeswithintheearcanalofasubject(Kidmoseetal., 2012).The rationaleforrecording fromwithintheear is unob-trusiveness (for hearingaid-related applications) and the hope thatmuscleartifactswillbeweaker,however theclosespacing betweenelectrodesisexpectedtoresultinsignalshighly corre-latedacrosschannels.Fig.4(a)showsashortsegmentofsignals fromthethreeelectrodes.Contrarytoexpectations,onechannel (green)isaffectedbychannel-specific glitchesof possible mus-cularorigin.ApplyingSTAR removestheseartifactsresultingin signalsthataremuchmoresimilaracrosschannels,asexpected fromcollocatedelectrodes(Fig.4(b)).Anotherwaytoattenuate channel-specificartifactsistoaverageacrosschannels,butthe ben-efithereisrelativelysmall(Fig.4(c),red).HoweverapplyingSTAR beforeaveraginggreatlyincreasesthisbenefit(blue).Thedeflection near0.5sisnotremovedbecauseitissharedacrosssensors.The effectofartifactreductionofthe3-channeldataisreflectedinits PCAspectrum(Fig.4(d)).Corticalsignalsfromcloselyspaced elec-trodesshouldbehighlycorrelated,implyingarapidlydecreasing PCAspectrum,andthisexpectation isbettermetafterapplying STAR.ThisexampleshowsthatSTARcanbeeffectiveevenfora smallnumberofchannels.

3.3. SimulatedtargetembeddedinrealEEG

This example involves a simulated repetitive target (half-sinusoidpulse)superimposedonthesame3-channelEEGsignal asinthepreviousexample.Theadvantageofusingasimulated targetisthatthegroundtruthisknown,theadvantageofusing realEEGasnoiseisthatitisrepresentativeofrealEEGnoise.The

(5)

0 0.5 1 1.5 2 s -200 0 200 V raw (a) 0 0.5 1 1.5 2 s STAR (b) 0 0.5 1 s -100 -50 0 50 100 V

average over channels

(c) raw STAR 1 2 3 PC 0.001 0.01 0.1 1 power (d) raw STAR 0 1 norm. amplitude raw (e) 0 1 norm. amplitude

cleaned with STAR (f) 0 0.5 1 s 0 1 norm. amplitude cleaned with JD (g) 0 0.5 1 s 0 1 norm. amplitude

cleaned with STAR + JD (h)

Fig.4.Three-channelEEG.(a)Rawdata.(b)SameafterartifactremovalbySTAR.(c) Averageoverchannelsofraw(red)orprocessed(blue)data.(d)Eigenvalue spec-trumofraw(red)andprocessed(blue)data.ThesmallerpowerofPCs2and3 indicatesthatthesignalsareclosertobeingidenticalacrosschannels,asexpected forthreecollocatedelectrodes.(e,f)Simulatedtarget(290repetitionsofa half-sinusoid)embeddedin3-channelEEGnoise(SNR=0.03).Blueisaverageovertrials, graybandindicates±2SDofabootstrapresamplingofthemean.(e)Averageover trialsofthe“best”electrode.(f)Sameas(e)afterapplyingtheSTARalgorithmto sup-pressartifacts.(g)Sameas(e)afterapplyingtheJDalgorithmtoenhancethetarget. (h)Sameas(e)afterapplyingfirstSTARthenJD.Thevaluesplottedarenormalized toequatethevarianceovertrialsineachcondition,thusthegreateramplitudeofthe mean(blue)in(h)reflectsthebetterSNRoftherecoveredsignal.(Forinterpretation ofthereferencestocolorinthislegend,thereaderisreferredtothewebversionof thearticle.)

targetpulsewasrepeated290timesandaddedtothesecond chan-neloftheEEGwithSNR=0.03.Fig.4(e)showstheaverageover trialsofthesecondEEGchannel.Thegraybandrepresents±2SD ofabootstrapresamplingofthemean.Thepresenceofartifactsis reflectedbytheirregularwidthofthegrayband.Inprinciple,the SNRofsuchmultichanneldatacanbeenhancedbytheJD algo-rithm(deCheveigné andParra,2014)withtheappropriatebias functiontoenhancerepeatablecomponents.Howeverinthiscase thebenefitislimited(Fig.4(g))presumablybecauseallchannels areartifact-contaminatedsothatthereisnoartifact-freesubspace. ApplyingtheSTARalgorithmattenuatestheseartifacts(Fig.4(f)) andJDproducesawell-isolatedtargetsignal(Fig.4(h)).This exam-pleshowsagainthatremovingchannel-specificartifactswithSTAR canimprovetheeffectivenessofcomponentanalysis.

4. Discussion

TheSTARalgorithmexploitsthebetween-channelcorrelation oftheunderlyingdatatoidentifychannel-specific artifactsand repairthem.Figuratively,it“surfs”thecorrelationstructureofthe underlyingdata,selectively shavingoffthepartsthatdonotfit. Channel-specificartifactsareubiquitousinEEGandotherrecording techniques.Removingthemisusuallytedious(ifdonemanually)

orcomplexandunreliable(ifdoneautomatically),andareliable automaticmethodtosuppressthemisthususeful.

4.1. Comparisonwithotherapproaches

Themostcommonapproachtochannel-specificartifactsisto discardtheoffendingchannel(s),orelsetheoffendingsamplesor trials(Pictonetal.,2000).Validdataarelostasaresult,andthe strategyfailsiftoomanychannelsand/orsamplesareaffected. Incontrast,STARentailsminimaldataloss,becauseonlyasmall proportionofsamplesarediscarded.Dataarerejectedona sample-by-samplebasis,incontrastforexampletoSCADS(Junghöferetal., 2000)thatremoveslargerchunks(trials)oftheaffectedchannel,or AFOP(Boudetetal.,2012)thatoperatesonarelativelylongwindow (e.g.20s).

Asinseveralotherapproaches,missingchannelsare interpo-latedspatially.InSTARtheinterpolationisbasedonanestimate ofthecorrelationstructuregatheredovertheartifact-freepart,in contrasttomethodssuchasSCADSthatapplysphericalspline inter-polationtoreconstructthemissingdata(Perrinetal.,1989).STAR thusrequiresnogeometricalinformation(itispurelydata-driven) althoughitmaybenefitfromusingaproximitymaptorestrictthe numberofchannelsuponwhicheachchannelisprojected.STAR usesthemultivariatecorrelationstructureofthedata,incontrast towaveletorempiricalmodebasis(Safieddineetal.,2012)thatuse thewithin-channelautocorrelationstructuretointerpolatemissing values.

Averydifferentapproachistousecomponentanalysistoform linearcombinationsofchannelsasinEq.(1),andprojectthe com-ponents that correspondto artifactsout of thedata. Thereare manytechniquesavailable(ICA,beamforming,JD,etc.)that dif-ferinthewaytheydeterminetheweightswjk.InICA(Delorme

etal.,2007;Maetal.,2012;Crespo-Garciaetal.,2008)ameasure ofstatisticalindependence(e.g.kurtosis)isusedtoconstrainthe weights.Channel-specificnoiseisusuallybothkurtotic(lotsofzero orextremevalues)andindependentbetweenchannels,sothis pro-cedureislikelytoisolatecomponentsthatcontainchannel-specific noise.Inbeamforming(Grosse-Wentrupetal.,2009)aspatialfilter isusedtoisolatesources,eitherdesired(toextractthem)or unde-sired(tosuppressthem),onthebasisoftheirspatialposition.InCSP (Kolesetal.,1990)orJD(deCheveignéandParra,2014)thespatial filterisinsteaddesignedtomaximizeorminimizesomedesirable orundesirabletraitofthedata.

Regardlessofthemethod,componentanalysisfacesa funda-mentallimitation:itcannotresolvemoresourcesthandimensions inthedata.STARisnotsubjecttothislimitation,asillustratedby theexampleinFig.1(d–g)whereartifactswerepresentonall chan-nels,inadditiontotargetandnoisesources,butthetargetcould nonethelessbeseparated.NeitherICA,norJD,norbeamformingcan dealwiththissituationbecausethenumberofsourcestoresolve exceedsthenumberofchannels.ThereasonwhySTARsucceedsis thatitoperateslocallyintime(sothatthesolutionappliedtofix eachsampleneednotconsiderothersamples),andleavesmostof thedataintact(thedimensionalityofthecleaneddataisthatof thedatabeforeartifactswereadded).Incontrast,techniquessuch asICAapplyasinglelineartransformtoallthedata,eachartifact removedreducesthedatadimensionality,andtheremaynotbe enoughdimensionstoallowthemalltoberemoved.Asstressed inSection1,anartifactspecifictochanneljcanonlyberemoved bysettingallthewjktozeroinEq.(2),i.e.discardingthat

elec-trode,andthereforeifartifactsarepresentonmanychannelsthe dimensionalityisgreatlyreduced.Empirically,theeffectivenessof ICAasa tooltoreduce myogenicartifactshasbeenquestioned because ICs may be found to contain both artifact and neural activity(McMenaminetal.,2010,2011),presumablyasa conse-quenceofthelargenumberofdistinctgeneratorsofmuscleartifact.

(6)

SimilartoSTAR,thesensornoisesuppression(SNS)method(de CheveignéandSimon,2008b)addressessensor-specificnoiseby replacingeachchannelbyitsprojectiononthesubspacespanned bytheotherchannels.HoweverwithSNSthisoperationisapplied uniformlyacrossthetimeaxis,andthussharesthesamedrawbacks ofothercomponentanalysistechniques.

TheAFOPmethodofBoudetetal.(2012)appliesdifferent solu-tionstodifferentpartsofthetimeaxis(ortime-frequencyplane). Artifacts withina limited spectrotemporalwindow areisolated using a variantof the commonspatial pattern (CSP) algorithm (Kolesetal.,1990),andthedataoverthatwindowarefixedby projectingthemout.AFOPcanremoveanyactivityspecifictothat window,notjustchannel-specific,incontrasttoSTARwhichonly removeschannel-specificartifacts.Howeverasnotedabove,AFOP operatesatacoarsertemporalgranularity.

ThefocusofSTARonchannel-specificartifactscouldbeseen asalimitation.Howeverthisrestrictionallowsittoavoid remov-ingactivitysharedacrosschannels,ofdeeperandpossiblyneural origin. Most importantly, STAR is compatible with other tech-niques(ICAorother)thatcandealwithchannel-sharedartifacts, increasing their effectiveness by relieving them of the burden ofchannel-specific artifacts.Indeed,theprimary motivationfor developingSTARwastopushthelimitsofthesenoisereduction techniques,includingourownpreviousefforts(deCheveignéand Simon,2007,2008a,b;deCheveigné,2010,2012;deCheveignéand Parra,2014).

AnfeatureofSTARisthatitisautomatic,incontrastto tradi-tionalartifactrejectionthatrequires manualintervention,orto ICAthat requiresinspectionof componentsto decidewhichto remove.Itisrelativelyparameter-free,incontrasttomethodssuch asFASTER(Nolanetal.,2010)thatinvolvemultipleparameters.It isalsoconceptuallysimpleandcomputationallyfast.

Tosummarize themainfeatures that distinguishSTARfrom otherapproaches:(a) itis automatic andrequires little tuning, (b) dimensionality is not reduced and apart from the artifact portionsthedataareuntouched,(c)themethodaddressesonly channel-specificartifacts,leavingotherformsofartifacttoother approaches,thatitcomplementsandcanenhance.

4.2. Caveatsandcautions

Assumption4requiresthatthedimensionalityoftheactivity underlyingthedata(e.g.numberofsignificantneuralsources)be lessthanthenumberofchannels.Astherearepotentiallybillions ofneuralsourcesthiscanonlybeanapproximation.

Thecorrelationstructure isassumedtobestationarysothat theprojectionparameters estimated fromthe artifact-free part canbeappliedintheartifact-contaminatedpart.Adeparturefrom thisassumption(forexampleifadditionalneuralsourcescoincided withmuscleartifact)wouldreduceeffectiveness.

Thecalculationoftheprojectionmatrix,thatallowseach chan-neltobereplacedbyitsprojectionontheotherchannels,issubject tooverfittingifthereisalargenumberofchannelsorinsufficient data.Itisusefultolimittheprojectiontoasubsetofneighboring channels(10seemstoworkwell).

Assumptions2and3requirethatchannel-specificartifactsbe localintime,which excludesslowchannel-specificfluctuations. Itmaybeusefultoremovesuchslowfluctuationsbyhigh-pass filteringthedatabeforeapplyingSTAR(orsplitthedatainto low-andhigh-passstreams,processthelatter,thenrecombine).

Foreachchannel,thealgorithmswitchesbetweenan artifact-free state, for which the data are untouched, and an artifact-contaminated state for which the data are replaced by their projection.Thetransitionbetweenstatesmayproduceastep,and thismayparadoxicallyleadtoincreasedhigh-frequencynoiseif thetransitionsarenumerousand/orthestepsarelarge.Smoothing

theeccentricitymeasurereducesthenumberofswitches,andin practiceswitchingseemsrarelyaproblemunlessthedatacontain alotoflow-frequencypower(seepreviouspoint).

Assumption2,thatartifactsshouldoccurononlyonechannel atatime,isoftenviolatedformyogenicactivitythatmayoccur inburststhataffectseveralchannelssimultaneously.Thishastwo consequences.First,thealgorithm(initsbasicform)onlyfixesone channelatagiventimepoint,andsoartifactsontheotherchannels mayremain.Second,projectiononartifact-contaminatedchannels maycontaminatethechannelbeingfixed.Theextensiondescribed inSection 2.4addressestheproblem,althoughitsapplicabilityis limitedbythecomputationalcostofdealingwiththemany com-binationsofartifact-contaminatedchannels(exponentialintheir number).TheNoiseToolstoolboximplementsanapproximation thatappearstobeeffective.STARmaybeappliedrepeatedly,so thateachnewpassremovesartifactsmissedbypreviouspasses. 4.3. Applicabilitytoelectromyogenicactivity

ThereisevidencethatmuchEEGpowerbeyond20Hzis mus-cularin origin: when muscular activityis silencedby injection of a paralysingdrug, powerin that bandis greatly diminished (Whithamet al.,2007).Thefact thatmuscularactivitycovaries withmentalstatemakesitapotentialconfoundincognitive stud-ies(Whithametal.,2008;Yuval-Greenbergetal.,2008;Shackman etal.,2009).TheSTARalgorithmoffersapartialsolution,inthatit removesthetypeofmyogenicactivitythatislocaltoone elec-trode (Ma et al., 2012).This constitutes a subset of myogenic activity:activitythatisinsteadcorrelatedacrosselectrodes(e.g. fromdeeperordistantmuscles)mustberemovedbyothermeans (McMenamin et al.,2010; Yilmaz et al.,2014).Empirically, the amountofreductionobservedforeachartifactburstisvariable, rangingfrom100%toafewtensofpercent(e.g.Fig.3).Artifactsthat arecorrelatedacrosselectrodes(e.g.fromocularordeepmuscles) maybeaddressedbycomponentanalysissuchasICAorJD,towhich STARiscomplementaryinthatitavoidssquanderingdimensions usefultothosemethods.TheSTARalgorithmisthusofpotential benefitinstudiesofcorticalactivityathigherfrequencies. 4.4. Applicabilitytoreal-timeandBCIapplications

Thealgorithmcanoperateonlineinrealtime.Afteraninitial bootstrapphase(togetanestimateoftheartifact-freecovariance matrix),thealgorithmswitchesbetweentwostates:(1)noartifact, thecovariancematrixisupdated,(2)artifact,theoffending chan-nelsarecorrected.Updatingofthecovariancematrixestimatemay occurforexampleaccordingtoaleakyintegratormechanism,to allowforslowvariationsofthecorrelationstructure.

ApplicationssuchasBCIinvolvingawearableEEGdeviceare limitedby artifactsrelated tomovement,includingcontactand muscleartifacts(Fatourechietal.,2007;McFarlandetal.,2005; Reisetal.,2014)thatcanforexampledisruptanotherwise effec-tivetechniquesuchasCSP(Grosse-Wentrupetal.,2009).Manual correctionorofflineprocessingareobviouslyruledout,andSTAR thusisofpotentialbenefitfortheseapplications.

5. Conclusions

Thesparsetimeartifactreduction(STAR)algorithmdeals effec-tively with channel-specific artifacts that are common in EEG, includingelectrodecontactnoiseandcertaintypesofmuscle arti-fact.Itoperateslocallyintime,correctingonlythesamplesaffected bytheartifact,andleavingtheremainderintact,andthushas min-imalnegativeimpactontheunderlyingdata.Removalofanartifact doesnotreducethedimensionalityofthedata,incontrastto chan-nelrejectionorcomponentanalysistechniquessuchasICA.The

(7)

STARalgorithm cancomplementthose techniquesand enhance theireffectivenessbyrelievingthemoftheburdenofdealingwith channelspecificartifacts.TheSTARalgorithmoperates automati-cally,requiresfewparameters,iscomputationallyefficient,andcan beimplementedtoperformonlineprocessinginrealtime.

Acknowledgements

ThisworkwassupportedbytheEUH2020-ICTgrant644732

(COCOHA),andgrantsANR-10-LABX-0087IECand

ANR-10-IDEX-0001-02 PSL*. Ed Lalor and Giovanni Liberto contributed the 128-channelEEGdata,andThomasLunnerandCarinaGraverson contributedthe3-channelEEGdata.

References

BoudetS,PeyrodieL,ForzyG,PintiA,ToumiH,GalloisP.Improvementsofadaptive filteringbyoptimalprojectiontofilterdifferentartifacttypesonlongduration EEGrecordings.ComputMethodsProgBiomed2012];108:234–49.

CorbyJC,RothWT,KopellBS.Prevalenceandmethodsofcontrolofthecephalicskin potentialEEGartifact.Psychophysiology1974];11:350–60.

Crespo-GarciaM,AtienzaM,CanteroJL. Muscleartifactremovalfrom human sleep EEG by using independent component analysis. Ann Biomed Eng 2008];36:467–75.

de CheveignéA. Time-shiftdenoising source separation.J Neurosci Methods 2010];189:113–20.

deCheveignéA.Quadraticcomponentanalysis.NeuroImage2012];59:3838–44. deCheveignéA,ParraLC.Jointdecorrelation,aversatiletoolformultichanneldata

analysis.NeuroImage2014];98:487–505.

deCheveignéA,SimonJZ.Denoisingbasedontime-shiftPCA.JNeurosciMethods 2007];165:297–305.

deCheveignéA,SimonJZ.Denoisingbasedonspatialfiltering.JNeurosciMethods 2008a];171:331–9.

de Cheveigné A, Simon JZ. Sensor noise suppression. J Neurosci Methods 2008b];168:195–202.

DelormeA,SejnowskiT,MakeigS.EnhanceddetectionofartifactsinEEGdata usinghigher-orderstatisticsandindependentcomponentanalysis.NeuroImage 2007];34:1443–9.

FatourechiM,BashashatiA,WardRK,BirchGE.EMGandEOGartifactsinbrain computerinterfacesystems:asurvey.ClinNeurophysiol2007];118:480–94. Goncharova II, McFarland DJ, Vaughan TM, Wolpaw JR. EMG

contamina-tion ofEEG: spectral andtopographical characteristics. ClinNeurophysiol 2003];114:1580–93.

Grosse-WentrupM,LiefholdC,GramannK,BussM.Beamforminginnoninvasive brain–computerinterfaces.IEEETransBiomedEng2009];56:1209–19. JunghöferM,ElbertT,TuckerDM.Statisticalcontrolofartifactsindensearray

EEG/MEGstudies.Psychophysiology2000];37:523–32.

KidmoseP,LooneyD,MandicDP.AuditoryevokedresponsesfromEar-EEG recor-dings.In:InternationalConferenceoftheIEEEEngineeringinMedicineand BiologySociety.Proceedings2012;2012].p.586–9.

KolesZ,LazarM,ZhouS.Spatialpatternsunderlyingpopulationdifferencesinthe backgroundEEG.BrainTopogr1990];2:275–84.

MaJ,TaoP,BayramS,SvetnikV.MuscleartifactsinmultichannelEEG:characteristics andreduction.ClinNeurophysiol2012];123:1676–86.

McFarlandDJ,SarnackiWA,VaughanTM,WolpawJR.Brain–computerinterface (BCI)operation:signalandnoiseduringearlytrainingsessions.Clin Neuro-physiol2005];116:56–62.

McMenaminBW,ShackmanAJ,MaxwellJS,BachhuberDRW,KoppenhaverAM, GreischarLL,DavidsonRJ.ValidationofICA-basedmyogenicartifactcorrection forscalpandsource-localizedEEG.NeuroImage2010];49:2416–32.

McMenaminBW,ShackmanAJ,GreischarLL,DavidsonRJ.Electromyogenicartifacts andelectroencephalographicinferencesrevisited.NeuroImage2011];54:4–9. MuthukumaraswamySD.High-frequencybrainactivityandmuscleartifactsin

MEG/EEG:areviewandrecommendations.FrontHumNeurosci2013];7:138. NiedermeyerE,LopezdaSilvaF.Electroencephalography:basicprinciples,clinical

applicationsandrelatedfields.LippincottWilliams&Wilkins;2005]. NolanH,WhelanR,ReillyRB.FASTER:fullyautomatedstatisticalthresholdingfor

EEGartifactrejection.JNeurosciMethods2010];192:152–62.

PerrinF, PernierJ, BertrandO, EchallierJF.Spherical splinesforscalp poten-tial and current density mapping. Electroencephalogr Clin Neurophysiol 1989];72:184–7.

PictonTW,BentinS,BergP,DonchinE,HillyardSA,JohnsonR,MillerGA,Ritter W,RuchkinDS,RuggMD,TaylorMJ.Guidelinesforusinghumanevent-related potentialstostudycognition:recordingstandardsandpublicationcriteria. Psy-chophysiology2000];37:127–52.

ReisPMR,HebenstreitF,GabsteigerF,vonTscharnerV,LochmannM. Methodologi-calaspectsofEEGandbodydynamicsmeasurementsduringmotion.FrontHum Neurosci2014];8:156.

SafieddineD, Kachenoura A, AlberaL, Birot G, KarfoulA,Pasnicu A,Biraben A,WendlingF,SenhadjiL,MerletI.RemovalofmuscleartifactfromEEG data:comparisonbetweenstochastic(ICAandCCA)anddeterministic(EMD andwavelet-based) approaches.Eurasip JAdv SignalProcess 2012];2012: 127.

ShackmanAJ,McMenaminBW,SlagterHA,MaxwellJS,GreischarLL,DavidsonRJ. Electromyogenicartifactsandelectroencephalographicinferences.BrainTopogr 2009];22:7–12.

WhithamEM,PopeKJ,FitzgibbonSP,LewisT,ClarkCR,LovelessS,BrobergM, Wal-laceA,DeLosAngelesD,LillieP,HardyA,FronskoR,PulbrookA,Willoughby JO.Scalp electrical recording during paralysis: quantitative evidence that EEGfrequenciesabove20Hz arecontaminatedbyEMG.ClinNeurophysiol 2007];118:1877–88.

WhithamEM,LewisT,PopeKJ,FitzgibbonSP,ClarkCR,LovelessS,DeLosAngelesD, WallaceAK,BrobergM,WilloughbyJO.ThinkingactivatesEMGinscalpelectrical recordings.ClinNeurophysiol2008];119:1166–75.

YilmazG,UnganP,SebikO,UginˇciusP,TürkerKS.Interferenceoftonicmuscle activityontheEEG:asinglemotorunitstudy.FrontHumNeurosci2014];8:504. Yuval-GreenbergS,TomerO,KerenAS,NelkenI,DeouellLY.Transientinduced gamma-bandresponseinEEGasamanifestationofminiaturesaccades.Neuron 2008];58:429–41.

Figure

Fig. 1. Denoising simulated data. (a) Target signal. (b) 10-channel data including target, Gaussian noise (SNR = 10), and an artifact on one channel
Fig. 3. 128-channel EEG data. (a) One channel with muscle artifact before (red) and after (blue) processing
Fig. 4. Three-channel EEG. (a) Raw data. (b) Same after artifact removal by STAR. (c) Average over channels of raw (red) or processed (blue) data

References

Related documents

About

The equilibrium involves specialized banks lending to one sector only. The proposition can be understood with the use of Figure 5. Consider a candidate equilibrium with two-

found that knockdown of HBXIP expression in HGC27-sh- HBXIP cells induced G0/G1 cell cycle arrest, whereas the up- regulation of HBXIP expression in SGC7901-LV-HBXIP cells had

Modulation of this type thus stands for a process where a particular English collocation forming part of the SL text is translated by two or more Slovene translation equivalents;

Based on the experimental result and the text files used it was concluded that RSA consume longest encryption time and AES algorithm consumes least encryption time.We also

If, following the Issue Date of Your Policy, Your Journey is cancelled, curtailed or unable to be completed because of the unforeseeable death, Accidental Injury, Sickness or

Be- cause the data obtained so far showed that (a) RNF146 translocates to the nu- cleus upon oxidative stress, (b) directly interacts with PARP-1 and (c) the level of both

Performance management practice in the public sector is specifically anchored around certain tools, drivers or antecedents, among which are performance measurement and