ContentslistsavailableatScienceDirect
Biomedical
Signal
Processing
and
Control
j ou rn a l h o m e p a g e :w w w . e l s e v i e r . c o m / l o c a t e / b s p c
Efficient
procedure
to
remove
ECG
from
sEMG
with
limited
deteriorations:
Extraction,
quasi-periodic
detection
and
cancellation
Franc¸
ois
Nougarou
a,∗,
Daniel
Massicotte
a,
Martin
Descarreaux
baDepartmentofElectricalandComputingEngineering,UniversityofQuébecatTrois-Rivières,Québec,Canada
bDepartmentofDepartmentofHumanKinetics,UniversityofQuébecatTrois-Rivières,Québec,Canada
a
r
t
i
c
l
e
i
n
f
o
Articlehistory:
Received1August2016
Receivedinrevisedform16June2017
Accepted20July2017
Availableonline4August2017
Keywords:
ECGcancellation
Quasi-periodicsignalsdetection
Discretewavelettransforms
Independentcomponentanalysis
a
b
s
t
r
a
c
t
Theinterpretationsofthesurfaceelectromyography(sEMG)signalsfromthetrunkregionarestrongly distortedbytheheartactivity(ECG),especiallyincaseoflow-amplitudeEMGresponsesanalyses.Many methodshavebeeninvestigatedtoresolvethisnontrivialproblem,byusingadvanceddataprocessing ontheoverallsEMGrecordedsignal.However,iftheyreduceECGartifacts,thosecancellationmethods alsodeterioratenoiselesspartsofthesignal.ThisworkproposesanoriginalECGcancellationmethod designedtolimitthedeteriorationofsEMGinformation.Todothat,theproposedtechniquesdoesnot directlyattempttoremovetheECG,butisbasedontwomainsteps:thelocalizationofECGandthe cancellationofECGbutonlywhereheartpulseshavebeendetected.Thephaseoflocalizationefficiently extractstheECGcontributionbycombiningthediscretewavelettransforms(DWT)andthemethodof independentcomponentanalysis(ICA).Andfinally,thisphasetakesadvantageofquasi-periodic prop-ertiesofECGsignalstoaccuratelydetectpulseslocalizationwithanoriginalalgorithmbasedonthe fastFouriertransform(FFT).Intensivesimulationswereachievedintermsofrelativeerrors,coherence andaccuracyfordifferentlevelsofECGinterference.Andthecorrelationcoefficientscomputedfrom theparaspinalmusclesEMGsignalsof12healthyparticipantswerealsousedtoevaluatethedeveloped method.Theresultsfromsimulationandrealdatademonstratethattheproposedmethodaccurately detectspulsespositionsandefficientlyremovestheECGfromEMGsignals,evenwhenbothsignalsare stronglyoverlapped,andgreatlylimitsthedeteriorationoftheEMG.
©2017ElsevierLtd.Allrightsreserved.
1. Introduction
Non-invasive and inexpensive, surface electromyography (sEMG)isaneffectivemethodthathasbeenwidelyadoptedin clini-calandresearchworkstostudytopicsrelatedtomuscularactivities suchasneuromusculardiseases.SurfaceEMGelectrodesprovidea non-stationaryelectricalsignalrepresentingthesumof subcuta-neousmotorunitactionpotentialsgeneratedduringamuscular contraction.However,beforeanyinterpretationsoftheobtained signals,post-processing isneeded. Indeed,sEMG electrodesare very sensitiveelements, and many artifact sources cancorrupt therecordedsignal:movementartifacts,interferencesfrompower lines,othersdevices,orothersbodyparts.Artifactscancellationis akeytopicinbiomedicalsignalprocessing.
Whenamuscularactivityisrecordedinthethoracicortrunk region,thesEMGsignalsarestronglycontaminatedbytheheart
∗ Correspondingauthor.
E-mailaddress:[email protected](F.Nougarou).
muscleselectricalactivity(ECG–electrocardiogram).Inthiscase, itisdifficulttoseparatetheheartcontributionfromthemuscular one,becausetheirfrequencyspectraoverlap:(i)from10to500Hz withmostpowerbetween20–200HzforEMGand(ii)from0to 100HzforECGwithapowerfailingafter35Hz[1].Cardiac arti-factsclearlycorruptmuscularresponses.Thiscorruptionismore damagingifthesignalstoanalysearelowmuscularactivities,as forexample,itisthecaseinstudiesabouttheimpactsofthespinal manipulationtherapyonthehumanphysiologicresponses[2–4]. Inthoseexperimentations,sEMGelectrodesrecordedspineerector spinaemuscles(closetothoracicvertebrae)obtainedduringspinal manipulationrealizedbyaservo-controlledmotor[2].ThosesEMG sensorsresponsesarelow-amplitudesignalsandarestrongly con-taminatedbycardiacpulses,becauseoftheirposition.Toaddress thissituation,thepresentpaperattemptstoproposeaneffective algorithm toremovetheECG,while maintainingthemaximum informationfromtargetedmuscles.
Duetothebiomedicalengineeringcommunityinterest,many methodstoremoveECGfromEMGsignalshavebeenproposedin theliteratureasexposedin[5,6].Simplehigh-passfiltersusedin http://dx.doi.org/10.1016/j.bspc.2017.07.019
[1,7]donotrepresentanacceptablesolutiontoreachthisstudy’s objective:EMGinformationistoomuchdeteriorated.Basedonan additionalelectrode,whichrecordsonlyECG,theadaptivenoise canceller(ANC)structurepresentsapowerfulsolutiontoremove ECG fromEMG signal.This methodhasbeen used from respi-ratoryand trapeziusmuscles in[8,9],bothwithRLS (Recursive LeastSquare)[10].Authorsin[11]proposeamodifiedversionof FBLMS(FastBlockLeastMeanSquares)toadjustthefilter coef-ficientsofANC,whichreturnsbetterperformancetocancelECG thanRLS.However,ANCperformancesaresensitiveto synchroni-sationbetweentheECGcontributionfrombothelectrodes(sEMG andECGreferenceelectrodes)andareoftenweakincaseofreal sig-nals.AstheANCmethod,theadaptivelineenhancer(ALE)[12]isa structurebasedonadaptivefiltering,whichdoesnotrequirea ref-erencesignaltoremoveECGfromEMGasin[13,14].Thistechnique allowstoefficientlyremovesignalnoiseandtoseparateperiodicor quasi-periodicnarrowbandcomponentsfromawidebandGaussian signal.Nevertheless,ALEcannotperformifthenoiseisnotwhite Gaussian[13],anditcanalsoremoveperiodicorquasi-periodic partsofdesiredEMGsignal.
IndependentComponentAnalysis(ICA)[15]andSingular Spec-trumAnalysis(SSA)[16],twomethodsofblindsourcesseparation, havebeenalsousedtosubstrateECGfromsEMGelectrodes.Based onseveralsEMGelectrodes,somemethodsuseICAtofirst decom-poseEMGsignalsintostatisticallyindependentcomponents.Then, the components containing ECG are discarded and remaining componentsareusedtoconstructcleanedsignals.In[17],ICA com-ponentscorrespondingtoECGweredeterminedusingtheperiodic characteristicsof cardiacpulses (RR interval).This processwas appliedontrunk,spinallumbaranddiaphragmmusclesin[17–21]. However,ECGandEMGcontributionsarenotperfectlyseparated withICA,and,whenECGcomponentsaretotallyremoved,EMG substantial values are lost. This deterioration is minimized by authorsin[15,22].Theyuseahigh-passfilterat30Hzcut-off fre-quencyonECGcomponentsbeforereconstructingsignals.Contrary toICA,basedonseveralsEMGelectrodes,SSAmethodusesseveral delayedversionsofasinglesignaltoextractindependent compo-nents[15].AtechniquebasedonSSAtodetermineandcancelthe ECGcontributionofthesignalwaspresentedin[14,23,24].It out-performstheALE[14],butsomeerrorsoccurwhentheperiodic signal(ECG)isnotnarrowbandorwelldefined[13].Inorderto overcomethisproblem,acombinationofSSAandALEwas sug-gestedin[13]andimprovedin[25],inwhichtheALE structure selectsadaptivelytheECGcomponentforwidebandperiodic sig-nalsdisturbedbynon-Gaussiannoise.AsclassicALE,thismethod candeteriorateEMGbyremovingitsperiodiccomponents.
Widelyappliedinbio-signalsprocessingsuchasevents detec-tion [26], classification [27] or artifacts cancellation, wavelet transform(WT)is a powerfultime-frequency approach[28,29]. However,WTdegradesEMGwhenit isdirectlyusedtoremove ECG.BetterperformancesareobtainedwhentheICAmethodis combinedtoWTinordertoselectandcancelECGcomponentsas proposedin[30–32].Thisstructurehasbeendesignedwithdiscrete andcontinuouswavelettransformin[33,34].
Thosetechniquesbased onICAand WT,like allother pow-erfulmethods exposed previously, apply theirtreatmentalong thewholesignal,evenduringnoiselesssections:theyefficiently removeECG,butgreatlydeterioratethedesiredEMG.The preser-vation of the EMG contribution during theECG cancellation is essential in this study. For this reason, the proposed method is based on a local cancellation approach. Contrary to other approaches,thismethoddoesnotdirectlycanceltheECG;itfirst focusesontheECG localization,beforeremovingit only where pulses positions have been estimated(cancellation phase). The majorcontributionisthelocalizationphasethatisableto accu-ratelylocatecardiacpulses,eveniftheECGandEMGamplitudes
areclose.First,theECGlocalizationrealizedbyefficiently combin-ingthediscretewavelettransform(DWT)andICAtoextracttheECG informationfromthesEMGelectrodes.Then,usingthisextracted information, an original pulses detection based onfast Fourier transform(FFT)takesadvantageofthequasi-periodicnatureofthe ECGtofindthecardiacpulsespositionswithastrongaccuracy.This stepoftheproposedmethodisessentialtogreatlyremoveECGwith limitedEMGdeteriorations.Indeed,asoftenwronglyassumedin literature,ICAdoesnotperfectlyseparateECGfromsEMG; partic-ularlyincaseofrealdataorwhenEMGandECGcontributionsare closed.
TheefficiencyofthecombinationofDWTandICAtofirstlocalize cardiacplusesandtheimpactofthetwophasesstructure (local-ization andcancellation) toremove ECGfromsEMG havebeen preliminarypresentedandevaluatedinaconferencepaper[30]. Inordertopreciselydescribethecompleteversionoftheproposed method(withadditionofthequasi-periodicECGpulsesdetection) andtorealizeanintensiveevaluation,thepresentpaperis orga-nizedasfollows:theSection1describesthesignalsmodelused forsimulation.Section2detailsthemainpartsoftheproposed methodforECGcancellation:theECGlocalizationwithextraction anddetectionofthecardiacsignal,followedbyitscancellation.A particularfocusontheproposedalgorithmofECGpulsesdetection basedonFFTisexposedinSection3.Theproposedmethodisthen evaluatedinSection4.Intensivesimulationshavebeenachievedto determinetheimpactofeachpartoftheproposedmethodinterms ofrelative errors (time-domain), coherence(frequency-domain) andaccuracyofitspulsesdetection.Simulationswereconducted fordifferentlevelsofECGinterferenceonEMG;theresultsobtained bymethodsbasedonclassicfilteringandSSAareusedfor com-parisons.Furthermore,inSection5,theperformancesofproposed methodareevaluatedwithrealdataobtainedduringthestudyof spinalmanipulationeffects[2–4].Finally,somebriefconclusions aredrawninSection6.
2. Signalmodel
Inthesignalmodel,realizedtoevaluateECGcancellation meth-ods,KfictiveclosedsEMGelectrodesareconsidered,whichrecord muscularactivitiescontaminated byECGfromthesamesource. Asdefinedin(1)andpresentedinFig.3a,sk[n] isthesumofthe
muscularcontributionsEMG
k [n] (information)andtheECG
compo-nentsECG
k [n] (interference)ofeachelectrode.Theindexesk=1,2,
...,Kandn=1,2,...,Narerespectivelyreferredtotheelectrodes andsamples,withNthesamplesnumberandFsthesampling
fre-quency.
sk[n]=sEMGk [n]+sECGk [n] (1)
TheinterferencelevelsarecontrolledwiththefollowingSIR (SignaltoInterferenceRatio)expressionindB,notedrSIR
k foragiven sEMGelectrodek: rkSIR=10log
Psk/PsECG k (2) wherePskisthepowerofthesignalsk[n] recordedbytheelectrodek,andPsECG k
thepowerofitsECGcontribution. ThesimulatedEMGcontributionsEMG
k [n] ofsk[n] areobtained
fromanautoregressivemodellingdescribedin[27].Foreach elec-trodek,thespectralcontentobtainedfromarealEMGwithout ECGisfirstinsertedinawhitenoisewithanormaldistribution. Then,theobtainedsignalsaremodulatedwithanenvelopeshape whichrepresentsrepeatedmusclecontractions(seeobtainedEMG signalsinFig.1a).ThesameenvelopeshapeisusedforallK elec-trodes.Basedonthisprocess,allsynthetizedEMGresponsesare closedbutdifferent.
Fig1.a)ExampleofobtainedEMGsimulatedsignalswithK=4electrodesfor
rSIR=13dB.b)ECGsimulatedsignals(focusedontwopulses)obtainedfrom(2)for
KelectrodesincomparisonwiththeoriginalrealECGsignal.
EvenifthesourceofECGartifactsisthesameforallKsEMG electrodes,allsECG
k [n] contributionsarenotexactlythesame,due
totheECGsignalpropagationintissuestoeachelectrode.TheK signalssECG
k [n] arecomputedfromthesamerealECGsignal,noted
¯sECG k [n],accordingtoexpression(3). sECGk [n]=g
¯sECGk [n]=k 1 1+exp(k+ ¯sECGk [n]) −1 (3) Thisexpressioninducesasmallnon-linearitytogenerateasmall deformationof ¯sECGk [n] toobtains ECG
k [n][11].g (•) isanon-linear
function,ksetsthedegreeofnon-linearityofthefunctionandk
controlstheamplitudeandthesignofthefunctionoutput.kand
karerandomlyfixedinalimitrangeinordertoinduce
signifi-cant,butsmalldeformationsbetweenallKelectrodesand ¯sECG k [n]
asshowninFig.1b.
3. ProposedmethodofECGcancellation
Theproposedmethod,presentedinFig.2,seekstominimizethe deteriorationoftheEMGcontributionduringtheECGremoval pro-cess.Todothat,thismethod,whichtakesadvantageoftheuseof severalsEMGelectrodes,isbasedontwomainphases:ECG
local-izationandECGcancellationbasedonknowledgeofcardiacpulses positions.
Detailsaboutmainstepsoftheproposedmethodaredescribed inthefollowingparagraphs.TheFig.3illustratesresultsobtained fromsimulateddataof4sEMGelectrodes.Inordertoclarifythe methoddescription,itisimportanttomakeadifferencebetween theterms“position”and“localization”.ThepositionofanECGpulse representsonesinglevalueofsample(itcanrefertothemaximum amplitudeofthepulse)andthelocalizationofanECGpulse corre-spondstothearea(agroupofsamples)wheretheoverallpulseis located.
3.1. ECGlocalization:DWT/ICAcombination
TheECGlocalizationfirstextractstheECGcontributionby per-formingacombinationoftheDWT[28]andtheICA.Asmentioned in[30],theDWTactsasfrequency/shapefilteringconfiguredto extractECGcomponentofeachrecordedsignal.Indeed,basedona given“motherwavelet” [n],DWTdecomposesallinputssk[n] in
severaltime-domainsub-signals,noteddjk[n],ofdifferent frequen-ciesranges,namedlevelsandreferredasj=1,2,...,J;withJthe totalnumberoflevels.Thefrequency/shapefilteringisthen per-formedbyselectingpertinentlevelsj.Becausethefrequencyrange ofmainECGcomponentsisbetween20Hzand50Hz,wavelet lev-elsofj=[4,5,6]areusedaccordingtoFs=1000Hz.Considering
[n] withashapeclosetocardiacpulse,thekthfrequency/shape
filteringresultxk[n] isasumoftime-seriesdjk[n] obtainedfrom
sk[n] forselectinglevels(4).Examplesofobtainedsignalsxk[n] are
exposedinFig.3b. xk[n]=
6 j=4d j k[n] (4)AsshowninFig.3aandb,thesignalsxk[n] containedessentially
ECGinformationbutalsosomeresiduesoftheEMGcontribution whichcanaltertheECGlocalization.ICAisappliedontheDWT frequency/shapefilteringresultsofallsEMGsensorstoimprove extractionofECGinformation.Indeed,asexposedin[22]and[30], theICAtechniquepermitstodecomposesomeinputsinto statis-ticallyindependentcomponentsfromeachother.Basedonxk[n],
theICAmethoddeterminessomeindependentsourcesyk[n] for
allk.AspresentedinFig.3c,theICAallowsregroupingthe com-moninformationfromthesignalsxk[n].The3rdICAcomponent
ofFig.3ccorrespondstoECGindependentcomponentwhichhas beenclearlyseparatedfromitsresidues.
3.2. ECGlocalization:quasi-periodicECGpulsesdetection
Based ontheICAindependentcomponent yk[n],the
“Quasi-periodicpulsesdetection”blockinFig.2appliesadetectionmethod takingadvantageofthequasi-periodicnatureoftheECGsignal.It returnsanestimationofthecardiacpulseslocalization,noted ˆ [n] (5).AsshowninFig.3d, ˆ [n] doesnotrepresentasinglesampleper pulse,butanestimatedareawherethepulsesarelocated.Details
Fig.3.IllustrationsoftheproposedmethodmainstepsforsimulatedsignalswithK=4sEMGelectrodes:a)KsynthetizedsEMGsignalswithECGforrSIR=19dB,b)Ksignals
filteredwiththeDWT,(6),c)KICAcomponents,(7),d)obtainedcardiacpulseslocalization,(8)(inblack)fromtheICAcomponentcorrespondingtoECG(inred),ande)K
EMGestimates,(9)(inblack)comparedtosignalswithECG(ingray).Infigures,amplituderangeisthesameforallelectrodesandtimerangeis0to10s.(Forinterpretation
ofthereferencestocolourinthisfigurelegend,thereaderisreferredtothewebversionofthisarticle.)
onthequasi-periodicECGpulsesdetectionprocessareexposedin theSection4.
ˆ [n] =
1, ifECGpulsesarepresent
0, else. (5)
3.3. ECGcancellation
ManytechniquesareavailabletocanceltheECGcontribution. Intheproposedmethod,Ksignalsk[n],whichrepresenttheECG,
areobtainedfromallsignalssk[n] byapplyingasimple4thorder
Butterworthlow-passfilterof50Hz.Asexposedin(6),thekthEMG
contributionsignal ˆsEMG
k [n] (Fig.3e)iscomputedbysubtractingthe
signalsk[n] ofthekthsEMGsensor(Fig.3a)withtheproductthe
signalk[n] andthevectorofECGlocalization ˆ [n].Forall
elec-trodesk,theECGcomponentissubtractedonlywherepulseshave beendetectedtolimitthedeteriorationofEMGinformation. ˆsEMG
k [n]=sk[n]−k[n] ˆ [n] (6)
4. Quasi-periodicECGpulsesdetectionmethod
EvenifthecombinationofwavelettransformandICAisefficient toextracttheECGcontributionfromasetofsEMGrecorded sig-nals,theautomaticlocalizationofECGpulsesremainsanon-trivial problem:somemistakesontheECGpositioncaneasilyappear, someartifacteventsbeingconsideredasECGpulses.Asmentioned inSection3.2,theproposedECGpulsedetectionmethodisbased onthequasi-periodicnatureoftheECGsignaltoestimateitspulses localization.Insteadofacommonlyusedamplitudethreshold,the proposedmethodusesaruleofperiodicitytodetermineECGpulses
positionstocreatetheECGpulseslocalizationsignal ˆ [n](5).Todo that,maineventspositions andperiodiceventspositionsofthe extractedECG contributionarecompared toestimate ˆ [n].The mainblocksfunctionsofthisprocedurearepresentedinFig.4and describedinthefollowingparagraphs.
4.1. ECGsignal&periodicfrequencyselection
Thefirststepoftheproposedpulseslocalizationmethod con-sistsinautomaticallyfindingys[n],theICAindependencesource
correspondingtotheECGcomponentfromyk[n],andits
funda-mentalfrequency,namedthefrequencyofperiodicityFp.TheECG
componentistheonewhichpresentsaclearperiodicity.The peri-odicityofasignalcanbeobservedinthefrequencydomain,not directlybyapplyingtheFFTonsignalsyk[n] forallK,butontheir
rectifiedversionsyrect
k [n],asdepictedin(7)andpresentedinFig.5a:
Yk[f ]=fft
yrect k [n] with yrect k [n]=|yk[n]|/max yk[n]|n=1,2,...,N (7)wherefft(䊉)isthefastFouriertransformfunction,|䊉|isthe abso-lutevaluefunctionandYk[f ],illustratedinFig.5.bforallK,isthe
magnitudeofthefrequencyrepresentationofthekthICArectified
componentyrect
k [n] withf=1,2,...,N.Notethatrectifiedsignals
havealsobeennormalizedbytheirmaximumvalue,withthe func-tion max(䊉), to allowa fair comparison betweenthe obtained frequencyresponses.InFig.5.b,themagnitudecorrespondingto theECGsignalpresentsthehighestfundamentalandharmonics peaks,whichattestthepresenceofperiodicity.Therectified trans-formation stands outthe periodicity of signals,since it reveals
Fig.4. Descriptionoftheproposedquasi-periodicpulsesdetectionmethodforand.
Fig.5. Illustrationofthe“ECGsignal&periodicfrequencyselection”blockwitha)K
rectifiedcomponentsfromICAprocessandb)magnitudesoffrequency
representa-tionsofsignalspresentedina).ThefrequencyrepresentationoftheICAcomponent
correspondingtoECGcontribution(the3thone)presentsaremarkable
fundamen-talfrequencynamedfrequencyofperiodicityandnotedFp.Infigures,amplitude
rangeisthesameforallelectrodes,timerangeis0to10sandfrequencyrangeis
0,5Hzto6Hz.
theirglobalbehavior.Inthispaper,theconsideredfrequencyrange of interestof the ECGsignal is between1Hz and 2Hhz, which correspondto60and 120pulsesperminute.Consequently,the index k of Yk[f ] which presents thehighest peak of frequency
(fundamentalfrequency)between1Hzand2Hz,correspondsto theECGsignalsandisnotedkECG.TherectifiedECGcomponent
isnotedys[n]=yrectk
ECG[n] andFp isthefundamentalfrequencyof
Yk
ECG[f ] (seeFig.5.b).
4.2. Maineventsdetection
TheobjectiveofthisblockpresentedinFig.4consistsinfinding themaineventspositionsˆe[n] from therectifiedECG
compo-nentys[n].Todothat,asliding-windowsmethod,wherevariance
iscomputedineachwindow,isappliedtodifferentiateimportant eventsofys[n] fromno-pertinentone.Theresultingsignalisthen
smoothedwithalowpassfilterandreferredas ¯ys[n] (seeingrayin
Fig.6d). ¯ys[n] allowsstandingoutmainpeakswhicharedetected
withapeaksdetectionmethodtoobtainˆe[n](8).Asshownin
Fig.6.d(backdottedlines),ˆe[n] containstheprecisepositionsof
ECGpulses,butalsoartifactpositions. ˆ
e[n]=
1, ifpositionof ¯ys[n] mainevent
0, else.
(8)
4.3. Periodiceventsdetection
In order toestimate quasi-pureperiodic eventspositions of ys[n],notedˆp[n](10),itisfirstnecessarytoconstructasignal
representingmainperiodiceventsofys[n],yps[n].Theideabehind
thatistoapplyonys[n] averyselectedband-passfiltercentered
onthefrequencyofperiodicity,Fp.However,becausestrong
ampli-tudesinfluencethefilteringresult,ys[n] istransformedbeforethe
filteringinordertoattenuatetheeffectofimportantevents.This transformation,basedonlogarithmfunctionlog (•) calledShannon entropyin[35],isdescribedinEq.(9).Comparedtoys[n] inFig.6a,
theresultingsignalylogs [n] whichappearedinFig.6bshowsthatall
peakshavenowthesameamplitudelevelandthattheinfluenceof extremepeaksiscancelled.
ylogs [n]=−|ys[n]|log (|ys[n]|) (9)
ylogs [n] isthenfilteredwithaveryselectedband-passfilter
cen-teredonthefrequencyofperiodicity,Fp,toobtainyps[n]Fig.6cin
gray.Periodicpeakspositionsofyps[n] arethendetectedandstored
inˆp[n] (blacklinesinFig.6c):
ˆ p[n]=
1, ifpositionofyps[n] mainevent
0, else.
(10) Contrarytoˆe[n],inˆp[n],thenumberofECGpulsespositions
isequaltothenumber ofpulsestodetect,but duetothestrict periodicrule,thosepositionsarenotpreciseasillustratedinFig.6c andd.
4.4. Quasi-periodiceventsestimation
Inthe“Quasi-periodiceventsestimation”block,finalestimated pulsespositions [n] areˆ obtainedbyusinganalgorithm,described inTable1,whichdeterminesgoodpositionsofˆe[n] (Fig.6d)in
functionofˆp[n] (Fig.6c).e[u] containsthesamplesvalues
cor-respondingtoˆe[n]=1andp[w] thesamplescorrespondingto
ˆ
p[n]=1,whereu=1,2,...,Neandw=1,2,...,Np.NeandNpare
respectivelythenumberofmaineventsandperiodicevents.For agivenw,thevectorofdistances,␦w,ofpositionp[w] fromthe
periodicpositionse[u] foralluiscomputedinlines2–4ofTable1
where␦w=
ıw[1] ,...ıw[u] ,...ıw[Ne]
T.Foragivenpositionw, ımin
w istheminimumdistanceof␦wforalluanduminw thisminimum
valueindex(seeline5inTable1).
Consideringˆanullvectorwithdim
ˆ=Np×1atthebegin-ningofthealgorithm,thefinalestimatedpositionsaredetermined for allindexwin functionof ımin
w asdescribedin lines6–10of
Table1.Fora givenw,iftheminimumımin
Fig.6.Illustrationsofmainstepsofthe“periodiceventsdetection”blockwitha)therectifiedECGcomponent,b)obtainedsignalafterlogarithmictransformation,c)obtained
filteredsignal(ingray)andperiodicpositions(blacklines).SmoothedsignalofrectifiedECGcomponentandrepresentationofobtainedmaineventspositions(blackdotted
lines)appearsind)topresentthe“Maineventsdetection”blockprocess.Thefinalestimationofquasi-periodicECGpositionsandobtainedECGpulseslocalizationare
respectivelyillustratedine)andf)(blacklines)withthesmoothedsignalofrectifiedECGcomponent(ingray).Timerangeoffiguresis0to10s.
Table1
Detectionalgorithmofquasi-periodicpositionsoftheECGsignal,[n]ˆ basedonmain
eventspositionse[n]ˆ andfromperiodiceventspositionsp[n]ˆ forn=1,2,...,N.
Inputs&Initialisation
p[w]:vectorofsamplescorrespondingtop[n]=1 e[w]:vectorofsamplescorrespondingtoe[n]=1 Fs&Fp:samplingfrequency&frequencyofperiodicity
ˆ = [0,0,...,0]Twithdim
ˆ=N p×1 Process 1 forw=1,2,...,Np 2 foru=1,2,...,Ne 3 ıw[u]=p[w]−e[u] 4 end 5 ımin w ,uminw =minıw[u]|u=1,2,...,Ne 6 ifımin w >Fs/2Fp 7 [ˆ p[w]]=1 8 else 9 [ˆ p[uminw ]]=1 10 end 11 end
orequaltohalfperiodicsample,Fs/2Fp,themaineventsposition ofthisminimumdistancee[uminw ]isconsidered‘good’.Ifthe mini-mumdistanceissuperiortoFs/2Fp,themaineventspositionofthe minimumdistanceistoofarfromtheperiodicityandisnot consid-ered‘good’.Inthiscase,theperiodicpositionoftheindexw,p[w], isconsidered.
4.5. Pulseswidthestimation
InordertoconstructthesignalofestimatedECGlocalization ˆl[n](seeFig.6f),itisfirstnecessarytoestimatethewidthofthe ECGpulses.ThisstepisbasedontheextractedECGcontribution
ys[n] andtheestimatedECGpositions [n].ˆ Asrealizedinthe“Main
events detection”block,ys[n] ismodified to obtaina smoothed
versionnoted ¯ys[n].Andthevectorofsamplescorrespondingto
ˆ
[n]=1,referredas [w] forw=1,2,...,Np,isalsoconsidered.
ForeachestimatedECGpulsew,athresholdwiscomputedfrom
thepeakvalueofthewthpulse ¯y
s[[w]],w=␣¯ys[[w]],where␣
isachosenpercentage(inthisproject␣=0.5%).Thenwisapplied
atleftandrightsideof [w],bykeepingvaluessuperiortowto
estimatethewidth [w] ofthiswthpulse.Inordertorejectthe
extremevaluesofdeterminedwidths,themedianvalueofwidth, med=median
[w]|w=1,2,...,Np
isconsidered.Thesignal ofECGpulseslocalization ˆ [n] isthenobtainedasaconvolutionof ECGpulsespositions [n] withˆ arectangularwindowofsize med. 5. Performanceevaluations
5.1. Simulationparameters,methodsandevaluationcriteria In order to evaluate the proposed method, K=4 sEMG electrodes were considered. According to (1) and (2), simu-lations were conducted for 9 levels of ECG contamination: SIR=[1,3,5,7,10,15,20,25,30]dBwherethesmallerSIRvalue repre-sentsastrongerimpactofECGontheEMGcontribution.Foreach SIRvalue,methodsofECGcancellationwereperformedNi=300
times.AsexposedinSection2,basedonrandomprocess,signals sk[n] forallk,allSIRvaluesandallsimulationrepetitionsare
dif-ferent,andtheirdurationisequalto10s(N=10000).
Theproposedmethod,referredasM3,usesinitsDWT filter-ingprocessamotherwavelets“db3”whichhasashapeclosetoa cardiacpulseandwaveletlevelsfixedinSection3.1.Forthesake ofcomparison,M3isevaluatedthroughsimulationiterationswith othertechniquesofECGremovallistedinTable2.M0showsresults obtainedwhennoECGremovemethodisused,inordertoobserve
Table2
SimulatedmethodstocancelECGandtheirdescription.
MethodsReferenceDescriptions
M0 Nomethodemployed
M1 Methodbasedona4thorderButterworthhigh-passat cut-offfrequencyof50Hz
M2 Methodbasedonsingularspectrumanalysis(SSA)
M3 ProposedmethodinFig.2,withDWT,ICAand
pseudo-periodicpulsesdetection
M4 MethodbasedonM3butwithoutpseudo-periodicpulses
detection
M5 MethodbasedonM3butwithoutICAandpseudo-periodic
pulsesdetection
M6 MethodbasedonM3withmoreaccuratecut-offfrequency
valuesinitshigh-passfilter
thelevelsofdegradationgeneratedbytheECGontheEMG sig-nalsforagivenSIRvalue.MethodsM1andM2,respectivelybased onclassicfiltering[1]andsingularspectrumanalysis(SSA)[16], directlyattempttoremovetheECGfromthesEMGwithout car-diacpulseslocalization.M2hasbeenchosenbecauseSSApresents anexcellentcapacitytoextractaperiodiccomponent(ECG)from agivensignal(EMG)[14]and[23].MethodsM4toM6represent particularcaseofM3.M4andM5allowobservingtheeffectof quasi-periodicpulsesdetectionandthecombination ofDWTand ICA, respectively.M4isthemethodpresentedintheconferencepaper [30].M6differsfromM3inchosencut-offfrequencyvaluesfc of
high-passfilterusedduringtheECGcancellation.M6takesinto accounttheamplitudeleveloftheECGcomparedtoEMG.Three levelsoftheECGtoEMGamplituderatioareconsidered:(i)ifthe ratioishighfc=50Hz,(ii)ifitismediumfc=45Hzand(iii)ifitis
lowfc=40HzM6allowsanalyzingtheimpactofamoreaccurate
techniqueinthe“ECGcancellation”blockofFig.2.
TherelativeerrorEkhasbeenchosentoevaluatetime-domain
similaritybetweentheEMGcontributionanditsestimate,sEMGk [n] and ˆsEMG
k [n].Smallerrelativeerrorvaluesreflectgreater
denois-ingperformances.Ekiscomputedasfollowforeachelectrodek,
method,andsimulationiteration: Ek=
N n=1 sEMGk [n]− ˆsEMG k [n] 2 /N n=1 sEMGk [n] 2 (11) Inaddition,a frequency-domaincriterion is providedbythe meancoherence[9],noted ¯Ck(12)forallelectrodesandmethodsateachiteration.Inordertofocusontherangeoffrequency contami-natedbytheECG,themeancoherenceiscomputedforfrequencies between10Hzand100Hz.Gooddenoisingperformanceresultsin coherencevaluesclosedto1.
¯Ck=mean
PsEMG k ˆs EMG k [f ] 2/P sEMG k s EMG k [f ] Pˆs EMG k ˆs EMG k [f ] (12) where PsEMG k ˆs EMG k [f ],PsEMG k s EMG k [f ] and PˆsEMG k ˆs EMG k[f ] are cross- and auto-spectral densities of sEMG
k [n] and ˆsEMGk [n], with 10Hz≤ f
≤100Hzasthefrequencysample.
Inordertoevaluatethecontributionofthequasi-periodicECG pulsesdetectionmethoddetailedinSection3,twootherscriteria havebeencomputedformethodsM3toM6ateachiteration:(i) errorsbetweenthenumberofECGpulsesdetectedbythemethods comparedtotherealnumberofECGpulses,εp,and(ii)themean
distancebetweenallestimatedpulsespositionsandtheircloser truepulsespositions,p.Amethodofpulsesdetectionisefficient
ofitisabletofindtherightnumberofpulseswithaclosedistance. Forallfourcriteria,meanshasbeencomputedalongall iter-ationsofsimulationandallelectrodes(Ni×K=300×4valuesper
mean)foragivenSIRvalue.Theobtainedcriteriameansinfunction oftheSIRvaluesforallmethodsarepresentedinFig.7forEk(11)
and ¯Ck(12),andinTable3formeannumberofpulseserrorεp,and
meandistanceerrorp.
Table3
Obtainedresultsofthemeannumberofpulseserrorandmeandistanceerrorin
functionofSIRvaluesforM3,M4andM5.
SIR(dB) Meannumberofpulseserror,εp Meandistanceerror(ms),p M3 M4 M5 M3 M4 M5 1 0.00 0.00 0.03 3.48 3.48 4.14 3 0.00 0.02 0.99 3.34 3.69 24.3 5 0.00 0.13 1.87 3.44 6.62 42.1 7 0.00 0.46 2.80 3.41 14.23 60.7 10 0.00 1.02 3.69 3.46 25.51 78.8 15 0.01 1.98 4.74 3.90 47.03 101 20 0.00 2.97 5.87 5.14 64.85 118 25 0.02 3.76 6.54 8.62 80.2 129 30 0.16 4.61 6.66 19.3 101 142
5.2. Evaluationresultsofsimulations
Theanalysisofeachmethodrequirestheinformationprovided byallcriteriacalculatedduringthesimulation.Forthisreason,the methodswillbeanalyzedonebyonefromtheresultspresentedin Fig.7aandb,andTable3(ifapplicable)startingwithmethodsM0, M1andM2:
• M0resultsin Fig.7 showthe damagecaused by theECG on EMGfortime-domain(Fig.7.a)andfrequency-domain(Fig.7b)in functionofthelevelofSIR.Inthosesimulations,theECGclearly interfereswiththeEMGsignal.
• Withitshigh-passfilterappliedtotheentiresignal,M1eliminates notonlytheECG,butalsotheEMG.InFig.7a,therelativeerror ofM1ishigherthanothermethods.Andobservingthefrequency domaininFig.7b,M1deterioratestheEMGsignalmorethanthe additionoftheECG(M0).
• M2failstofinelyseparateECGcomponentofEMG:itsresultinthe frequency-domaincannotdobetterthanM0.AstheECGsignal, thesynthetizedEMGpartofallelectrodes(Fig.1a)alsopresenta periodicity,withrepetitivemuscularcontractions.Thisfact dis-turbstheSSAmethod,whichtrytofindperiodicindependent componentsofagivensignal.
Belowisathoroughanalysisoftheproposedmethod,M3. Meth-odsM4andM5,whichareincompleteversionsofM3areusedhere tounderstandtheimpactofvariouspartsofM3.
• AsobservedinFig.7a,M3allowstoclearlyobtainsmallerrelative errorsthanM0,M1andM2,whatevertheappliedECGlevel(SIR). Fig.7bshowsthatM3alsoimprovesthefrequencycontentof thedegradedsignal:forallSIR,theobtainedfrequencycontentis closertotheoneofthesignalwithoutdeterioration.M3efficiently cancelsthecardiacpulsesfromtheEMGsignal.TheM3results comparedtoM1andM2demonstratethevalueofitstwo-steps strategytoremoveECG:localizationandcancellation.
• ResultsinTable3showthatthedetectionofthecardiac pulsa-tionusedinM3isverypowerful.WhatevertheECGdeterioration level,theerrorofdetectedpulses,εp,islessthan0.16pulses(for
SIR=30dB). Furthermore,thepositionsofthedetectedpulses inM3areclosetothepositionsofaddedpulsestoEMG: max-imumaveragedistancebetweenthedetectedandaddedpulses isp=19.3samples(SIR=30dB),whichisequivalentto19.3ms.
EveniftheeffectoftheheartbeatissmallcomparedtotheEMG, M3allowstoaccuratelydetecttheECGpulses.
• ContrarytoM3,M4doesnottakeadvantageoftheperiodicnature oftheECGsignalintheirpulsesdetectionstep.AsshowninFig.7a, M4providesgood relativeerrorresults:lessthanM0, regard-lessoftheinterferenceslevel.However,inFig.7b,M4improves thefrequencycontentoftheinterferedsignal(greaterthanM0) only untilSIR=20dB. ComparingM4 toM3,resultsare equal
Fig.7.ObtainedresultsforallECGremovemethodsof(a)meanrelativeand(b)meancoherenceinfunctionofSIRvaluesindB.
until7dB, but M4resultsdeteriorateafter this SIRlevel. The quasi-periodicpulsesdetectionofSection3improvestheECG cancellationresultsofM3.However,M3andM4returnbetter resultstoremoveECGfromEMGthanmethodsbasedonICA(as methodproposedin[17])whichrejecttheentireICAcomponents correspondingtoECG.Indeed,contrarytothosemethods,M3and M4,withtheirtwomainphasespreservenoiselesspartsofthe signal.
• ThecomparisonofM4andM5allowsevaluatingthecombination ofDWTandICA:M4usesthiscombination[30],whileM5only usesDWTtoextracttheECGcomponentfromtheEMGsignal priortodetectionpulses(Fig.2).WithoutICA,M5returnsworse resultsin termsof relative errorthanM4.Theinterest ofthe combinationespeciallyappearsinFig.7bwherethefrequency contentobtainedwithM4isclearlyclosertothefrequency con-tentofthesignalwithoutinterferencethanM5.Furthermore, Table3showsthatM5hasmoreerrorsinthedetectedpulses numberandinaccuracythanM4forallSIRlevels.
• InTable3,εp andpincrease forM4andM5whenthelevel
ofinterferencesdecreases.Indeed,itismoredifficulttodetect heartbeatswhentheirinfluenceisweak.Howevertheerroron thenumberofpulsesremainsalmostconstantandcloseto0for M3.Furthermore,theworstprecisionerrorobtainedforM3at SIR=30dBisachievedapproximatelybyM4atSIR=10dBand M5atSIR=3dB.ThecombinationofDWT,ICAandthedetection basedonthequasi-periodicsignaleventsusedinM3efficiently localizestheECGpulses,evenwhentheyaresmallcomparedto EMG.
Finally,oncetheECGpulsesarelocalized,severaltypesof can-cellationmethodscanbeused.MethodsM3andM6areidentical exceptfortheir“ECGcancellation”blockcontent.M6,exposedin Section4.1,minimizesdamagetotheEMGsignalwhentheimpact oftheECGislower.RegardlessoftheSIRvalues,relative errors andmeancoherencesareimprovedwithM6comparedtoM3.The proposedmethod,whenthepulsesarecorrectlylocalizedandthe cancellationmethodiswelldesigned,canefficientlyremovethe ECGpulsesfromtheEMGsignal,evenwhentheyaremoredifficult todetect.
5.3. Experimentalvalidationonrealsubjects
SomeexperimentationsareconductedattheGRANlaboratory (chiropracticdepartment,UniversityofQuebecatTrois-Rivières, Qc, CA) to study fundamental aspects the spinal manipulation therapy(SMT)[3,4].Themanipulationwasperformedbya servo-controlledmotor[2] onthe7th vertebraethoracic (T7)(see an exampleofappliedforcecurveinFig.8firstgraph).Bipolar muscu-larresponseswererecordedaroundtheimpactwithfourDelsys
Surface EMG electrodes (Model DE2.1, DelsysInc., Boston, MA, USA).Thosesensorswerepositionedoneachsideofthespineover thethoracic spineerectorspinaemuscles:two sEMGsensorsat 2cmfromthe6thvertebraethoracic(T6)andtwoothersfromthe 8thvertebraethoracic(T8).DatawererecordedwithFs=1000Hz.
BeforeapplyingtheECGcancellationmethods,a10–500Hz band-pass 4th order Butterworth filterwas appliedin order toonly considertheEMGinformation[1].Thepowerlineinterferencewas cancelledwithnotchfiltersat60Hzanditsharmonics.
Thebehavioroftheproposedmethod(M3) withrealdatais observedinFig.8wheresEMGelectrodealsorecordscardiacpulses with different shapes. M3 efficiently detects and removes car-diacpulses,leavingintactnotinterferingpartsofthesignals.This methodlimitsthedamagetothesignalofinterestandtheerrors ofsignalsinterpretation.
InordertoevaluateM3withrealdata,twelvehealthyvolunteers withmeanageof23.52±3.1yearsweresubjectedtofourSMTforce curveswithdifferentpeakforcesof65,100,150and225N. Partic-ipantsandappliedforcesarereferredtoo=1,2,...,12andi=1,2, ...,4.Thecorrelationcoefficientso,i,k(13)ofallsubjectso,allSMT
forcecurvesiandallelectrodeskarecomputedtocompare perfor-mancesofM1,M2,M3andM4.Thisquantitativeassessment,used in[13],islowwhenestimatedEMG ˆsEMG
o,i,k[n] andECGparts ˆs ECG o,i,k[n]
areuncorrelated,orsupposedtobeclearlyseparated.
o,i,k= 1 N
N n=1ˆs EMGo,i,k[n] ˆsECGo,i,k[n] (13)
Fig.9.apresentsmeansofthecorrelationcoefficientsforall par-ticipantsandappliedforcesandshowsthatM3reducesECGmore accuratelythanothermethods.Indeed,aone-wayrepeatedANOVA wasperformedonobtainedcorrelationcoefficientsandreportsa significanteffectonmethods
[F(3,1161)=16.319,p=0.000025].Then,aTurkeypost-hoctest revealsthatM3presentscorrelationcoefficientssignificantlylower thanM1,M2andM4(p=0.00008).Indeed,M1andM2cancel noise-lesspartsofsignal.IdenticaltoM3butwithalesseffectivepulses detectionmethod,M4returnsasimilarresultthanM1andM2.The performanceofM4canbeexplainedwiththefigureFig.9bwhere mean ofcorrelation coefficients are presentedfor each applied peakforce.Contrarytoothermethods,coefficientsobtainedfor M4increasewiththelevelofpeakforce.Asexplainedinthestudy exposedin[3],thehighertheappliedpeakforceisandthehigheris themuscularactivityresponse.Theapplicationofquickpeakforce createsmuscularbursts(EMG)whichisidentifiedbythedetection methodofM4asECGpulsations.EvenifsignalspresentsomeEMG bursts,theproposedmethodM3preservestheEMGinformation andrejectsECGpulsations.
Fig.8. ObtainedresultsaftertheuseoftheproposedECGcancellationmethodonsEMGelectrodessignalsfromrightandleftsideofT6andT8.Measuredforceappliedon
spineatT7appearsinthefirstfigure.
Fig.9. ObtainedmeanandstandarddeviationofcorrelationcoefficientsformethodsM1,M2,M3andM4(a)withnodistinctionofappliedforcesand(b)foreachlevelof
appliedpeakforces.
6. Conclusion
Muscular activities(EMG) recorded in thethoracic or trunk regionarestronglycontaminatedbythecardiacpulsations(ECG), whichlimittheinterpretationsoftheEMGsignals,especiallyincase oflow-amplituderesponses.ECGartifactsmustberejectedandthe EMGpreservedasmuchaspossible.Forthisreason,theproposed methoddoesnotdirectlyattempttoremovetheECG,butactsintwo mainphases:thelocalizationoftheECGpulsesandtheir cancella-tiononlywheretheyhavebeenlocated.Thelocalizationphaseuses acombinationofDWTandICAtoefficientlyextractECG,andthen takesadvantageofitsquasi-periodicnaturetoaccuratelydetect pulseswithaproposedmethodbasedonFFT.
Thesimulations,conductedintime-andfrequency-domainfor differentlevels of ECG contamination,and theevaluation from experimentaldata demonstratethat theproposedmethod effi-cientlycancelstheECGpulsesfromtheEMGsignal,evenwhen theyaremoredifficulttodetect,andlimitsdamagesontheEMGin thenoiselesspartsofthesignal.Inaddition,thebeneficialimpact ofthelocalECGcancellation,thecombinationofDWTandICA,and theproposedpulsesdetectionhaveclearlybeenshown.
Severalresearchperspectivescanbeconsideredwiththis pro-posedmethod.Actuallyperformedbyasimplehighpassfilter,ECG cancellationphase couldbeimproved. Pulseslocalization could beefficientlyusedinanadaptivestructure,orwithothers tech-niquestorealizelocalrejection.Note that,theoverall structure ofthismethodisadequatetoapplyaniterativeprocesstorefine thecancellationperformancesateachiteration.Finally,the
pro-posedmethodwithitsefficientquasi-periodicsignalsdetection couldbeintroducedinothercontextcharacterizedbythepresence ofperiodicorquasi-signalsforcancellationordetection.
Acknowledgements
ThisworkissupportedinpartbytheFondsExcellenceUQTRand theNaturalSciencesandEngineeringResearchCouncilofCanada (NSERC).
References
[1]J.D.Drake,J.P.Callaghan,Eliminationofelectrocardiogramcontamination fromelectromyogramsignals:anevaluationofcurrentlyusedremoval techniques,J.Electromyogr.Kinesiol.16(2)(2006)175–183(2005).
[2]M.Descarreaux,F.Nougarou,C.Dugas,Standardizationofspinal
manipulationtherapyinhumans:developmentofanoveldevicedesignedto measuredose-response,J.ManipulativePhysiol.Ther.36(2)(2013)78–83.
[3]F.Nougarou,C.Deslauriers,C.Dugas,M.Descarreaux,Physiologicalresponses tospinalmanipulationtherapy:investigationoftherelationshipbetween electromyographicresponsesandpeakforce,J.ManipulativePhysiol.Ther.39 (9)(2013)557–563.
[4]F.Nougarou,C.Dugas,M.Loranger,I.Pagé,M.Descarreaux,Theroleof preloadforcesinspinalmanipulation:experimentalinvestigationof kinematicandelectromyographicresponsesinhealthyadults,J.Manipulative Physiol.Ther.37(5)(2014)287–293.
[5]B.S.Darak,S.M.Hambrade,Areviewoftechniquesforextractionofcardiac artefactsinsurfaceEMGsignalsandresultsforsimulationofECG-EMG mixturesignal,in:IEEEInternationalConferenceonPervasiveComputing (ICPC),April2015,2015,pp.1–5.
[6]N.Miljkovi ´c,N.Popovi ´c,O.Djordjevi ´c,L.Konstantinovi ´c,T.B. ˇSekara,ECG artifactcancellationinsurfaceEMGsignalsbyfractionalordercalculus application,J.Comput.MethodsProgramsBiomed.140(2017)256–264.
[7]M.S.Redfern,R.E.Hughes,D.B.Chaffin,High-passfilteringtoremove electrocardiographicinterferencefromtorsoEMGrecordings,Clin.Biomech. 8(1)(1993)44–48.
[8]M.Golabbakhsh,M.Masoumzadeh,M.FarzanSabahi,ECGandpowerline noiseremovalfromrespiratoryEMGsignalusingadaptivefilters,MajlesiJ. Electr.Eng.5(4)(2011)28–33.
[9]G.Lu,S.Brittain,P.Hollan,J.Yianni,A.L.Green,F.Stein,T.Z.Aziz,S.Wang, RemovingECGnoisefromsurfaceEMGsignalsusingadaptivefiltering, Neurosci.Lett.462(1)(2009)14–19.
[10]S.Haykin,AdaptiveFilterTheory,4thed.,PrenticeHall,2002.
[11]M.E.F.Djellatou,F.Nougarou,D.Massicotte,EnhancedFBLMSalgorithmfor ECGandnoiseremovalfromsEMGsignals,in:IEEE18thInternational ConferenceonDigitalSignalProcessing(DSP),1–6,July,2013.
[12]B.Widrow,J.R.Glover,Jr.,J.M.McCool,J.Kaunitz,C.S.Williams,R.H.Hearn, etal.,Adaptivenoisecancelling:principlesandapplications,Proc.IEEE63 (12)(1975)1692–1716.
[13]S.Sanei,T.K.M.Lee,V.Abolghasemi,Anewadaptivelineenhancerbasedon singularspectrumanalysis,IEEETrans.Biomed.Eng.59(February(2))(2012) 428–434.
[14]S.Sanei,A.R.Hosseini-Yazdi,ExtractionofECGfromsinglechannelEMG signalusingconstrainedsingularspectrumanalysis,in:IEEE17th InternationalConferenceonDigitalSignalProcessing(DSP),1–4,July2011, 2011.
[15]Y.Hu,X.Li,X.Xie,L.Pang,Y.Cao,K.Luk,Applyingindependentcomponent analysisonECGcancellationtechniqueforthesurfacerecordingoftrunk electromyography,AnnualInternationalConferenceofIEEEEngineeringin MedicineandBiologySociety(2005)3647–3649.
[16]N.Golyandina,V.Nekrutkin,A.Zhigljavsky,AnalysisofTimeSeriesStructure: SSAandRelatedTechniques,Chapman&Hall/CRC,NewYork,2001.
[17]J.N.F.Mak,Y.Hu,K.D.K.Luk,AnautomatedECG-artifactmethodfortrunk musclesurfaceEMGrecordings,J.Med.Eng.Phys.32(2010)840–848.
[18]J.N.F.Mak,Y.Hu,K.Luk,ICA-basedECGremovalfromsurface
electromyographyanditseffectonlowbackpainassessment,IEEEEMBS Conf.NeuralEng.(2007)646–649.
[19]J.D.Costajr,D.D.Ferreira,A.M.F.L.MirandadeSá,Reducing
electrocardiographicartifactsfromelectromyogramsignalswithIndependent ComponentAnalysis,in:IEEEEMBSInternationalConference,September 2010,2017,pp.4598–4601.
[20]Y.Hu,J.N.F.Mak,K.D.K.Luk,Effectofelectrocardiographiccontaminationon surfaceelectromyographyassessmentofbackmuscles,J.Electromyogr. Kinesiol.19(1)(2009)145–156.
[21]S.R.Alty,W.D.C.Man,J.Moxham,K.C.Lee,Denoisingofdiaphragmsignalsfor respiratorycontrolanddiagnosticpurposes,AnnualInternationalConference ofIEEEEngineeringinMedicineandBiologySociety(2008)5560–5563.
[22]Y.Hu,J.Mak,H.Liu,K.D.K.Luk,ECGcancellationforsurface
electromyographymeasurementusingindependentcomponentanalysis, IEEEInt.ISCAS(2007)3235–3238.
[23]S.Sanei,M.Ghodsi,H.Hassani,Anadaptivesingularspectrumanalysis approachtomurmurdetectionfromheartsounds,J.Med.Eng.Phys.33(3) (2011)362–367.
[24]J.Barrios-Muriel,F.Romero,F.J.Alonso,K.Gianikellis,AsimpleSSA-based de-noisingtechniquetoremoveECGinterferenceinEMGsignals,J.Biomed. SignalProcess.Control30(2016)1117–1126.
[25]M.E.F.Djellatou,D.Massicotte,M.Boukadoum,AdaptiveblockSSAbasedANC implementationforhighperformancesECGremovalfromsEMGsignals,in: IEEE27thCanadianConferenceonElectricalandComputerEngineering (CCECE),May2014,2014,pp.1–6.
[26]F.Nougarou,D.Massicotte,M.Descarreaux,Detectionmethodof
flexion-relaxationphenomenonbasedonwaveletsforpatientswithlowback pain,EURASIPJ.Adv.SignalProcess.(2012)1–17.
[27]A.M.Subasi.Yilmaz,H.R.Ozcalik,ClassificationofEMGsignalusingwavelet neural,J.Neurosci.Methods156(September(1–2))(2006)360–367.
[28]S.Mallat,AWaveletTourofSignalProcessing,AcademicPress,1998.
[29]A.Medl,Timefrequencyandwaveletsinbiomedicalsignalprocessing,IEEE Eng.Med.Biol.Mag.17(December(6))(1998).
[30]F.Nougarou,D.Massicotte,M.Descarreaux,EfficientcombinationofDWTand ICAtolocalizeandremoveECGfromsurfaceelectromyography
measurement,in:IEEEInternationalConferenceonDigitalSignalProcessing (DSP),July,2013.
[31]S.Abbaspour,M.Linden,H.Gholamhosseini,ECGartifactremovalfrom surfaceEMGsignalusinganautomatedmethodbasedonwavelet-ICA,Stud. HealthTechnol.Inf.211(2015)91–97.
[32]S.Abbaspour,A.Fallah,M.Linden,H.Gholamhosseini,Anovelapproachfor removingECGinterferencesfromsurfaceEMGsignalsusingacombined ANFISandwavelet,J.Electromyogr.Kinesiol.26(2016)52–59.
[33]J.Taelman,S.VanHuffel,A.Spaepen,Wavelet-independentcomponent analysistoremoveelectro-cardiographycontaminationinsurface electromyography,AnnualInternationalConferenceofIEEEEngineeringin MedicineandBiologySociety(2007)682–685.
[34]V.vonTscharner,B.Eskofier,P.Federolf,Removaloftheelectrocardiogram signalfromsurfaceEMGrecordingsusingnonlinearlyscaledwavelets,J. Electromyogr.Kinesiol.21(4)(2011)683–688.
[35]S.S.Manikandan,K.P.Soman,AnovelmethodfordetectingRpeaksin electrocardiogram,J.Biomed.SignalProcess.Control7(2)(2011)118–128.