Pleasecitethisarticleinpressas:M.Meo,etal.,Catheterablationoutcomepredictioninpersistentatrialfibrillationusingweightedprincipal ContentslistsavailableatSciVerseScienceDirect
Biomedical Signal Processing and Control
j o u r n al ho 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
Catheter ablation outcome prediction in persistent atrial fibrillation using weighted principal component analysis
Marianna Meo
a,∗, Vicente Zarzoso
a, Olivier Meste
a, Decebal G. Latcu
b, Nadir Saoudi
baLaboratoired’Informatique,SignauxetSystèmesdeSophiaAntipolis(I3S),UniversitéNiceSophiaAntipolis,CNRS,France
bServicedeCardiologie,CentreHospitalierPrincesseGrace,Monaco
a r t i c l e i n f o
Articlehistory:
Received22June2012
Receivedinrevisedform1February2013 Accepted11February2013
Available online xxx Keywords:
Atrialfibrillation(AF) Catheterablation(CA) Electrocardiogram(ECG) Spatialdiversity
Weightedprincipalcomponentanalysis (WPCA)
a b s t r a c t
Radiofrequencycatheterablation(CA)isincreasinglyemployedtotreatpersistentatrialfibrillation(AF), yetselectionofpatientswhowouldpositivelyrespondtothistherapyiscurrentlyacriticalproblem.Sev- eralparametersofthesurface12-leadelectrocardiogram(ECG)havebeenanalyzedinpreviousworksto predictAFterminationbyCA.Nevertheless,theyareaffectedbysomelimitations,suchasmanualcom- putationandtheexaminationofasingleECGleadwhileneglectingcontributionsfromotherelectrodes.
AFspatio-temporalorganizationhasbeendescribedonsurfaceECGbymeansofthenormalizedmean squareerror(NMSE)betweenconsecutiveatrialactivity(AA)signalsegmentsandtheirreduced-rank approximationsbasedonprincipalcomponentanalysis(PCA).However,thesefeaturesdonotshowto becorrelatedwithCAoutcome.Inthisstudy,suchdescriptorsareadequatelyadaptedandappliedtoCA outcomeprediction.AnNMSEindexisputforward,computedoverthesetofeightlinearlyindependent ECGleadsafterAAsignalrank-1approximationsdeterminedbyweightedprincipalcomponentanal- ysis(WPCA).Thefinalpredictorisabletodiscriminatebetweensuccessful(70.76±17.74)andfailing CAprocedures(37.54±20.01)beforeperformingtheablation(p-value=0.0013,AUC=0.91).Thepro- posedWPCA-basedtechniqueemphasizesthemostdescriptivecomponentsofAFelectrophysiology byselectivelyenhancingcontributionscomingfromthemostrepresentativeECGleads.Ourinvestiga- tionconfirmsthatECGspatialdiversityexploitationinthisWPCA-basedframeworknotonlyendowsthe NMSEindexwithclinicalvalueinthecontextofCAoutcomeprediction,butitalsoimprovesclassification accuracyandincreasesrobustnesstoECGleadselection.
© 2013 Elsevier Ltd. All rights reserved.
1. Introduction
Atrialfibrillation(AF)isasustainedcardiacarrhythmiachar- acterizedbyrapidanddisorganizedatrialactivationsinducinga lossofatrialmechanicalefficacy.Severaltheorieshavebeensug- gestedtoexplainAFelectrophysiologicalmechanisms,soastoput forthasystematicproceduralprotocolforitstreatment.AFactiv- ityhasbeenfirstregardedastheresultofinteractions between multiplewanderingatrialwavelets[1,2].Ontheotherhand,itis commonlyacknowledgedthatpulmonaryveins(PVs)significantly contributetoAFmaintenanceandevolution,especiallyinparox- ysmalformsofthisdisease[3].Inspiteofmajoradvancesinits treatment,AFremainsasignificantcauseofcardiovascularmor- bidityandmortality,especiallythosearisingfromstrokeandheart failure.
∗Correspondingauthor.Tel.:+33492942741.
E-mailaddresses:[email protected](M.Meo),[email protected](V.Zarzoso), [email protected](O.Meste),[email protected](D.G.Latcu),[email protected] (N.Saoudi).
Radiofrequencycatheter ablation (CA)hasbecome thefirst- linestrategy [4] for thetreatment of this disease.However,as theprecisepathophysiologyof AFdynamicshasnotbeencom- pletelyclarifiedyet,itisstillquestionablewhetherCAeffectively suppressesabnormalrhythmsources,andhowitaffectsheartelec- tricalsubstrate.DifferentCAtechniqueshavebeendeveloped,yet noneofthemiswidelyconsideredaseffectiveforthetreatmentof persistentAF.Theirperformanceisstillfarfromsatisfactory,and theyarelesseffectivethanequivalentproceduresforparoxysmal AF.Sincethiscardiacinterventionalprocedureisprofoundlyinflu- enced byoperator’sexperience and patient’shealth conditions, resultsreportedbyclinicalcentersarequitedisparateandnoteas- ilycomparable[5–7].ItfollowsthatitsefficacyinterminatingAF andavoidingitsrecurrenceisnotguaranteedforallpatients.This situationexplainstheincreasingtendencytoattemptanapriori selectionofpatientswhocanundergoCAandexperiencedurable sinusrhythm(SR)restoration.Severalparametersextractedfrom thesurfaceECGhavebeenproposedaspotentialpredictorsofCA outcome[8,9].Forexample,prolongationofatrialfibrillationcycle length(AFCL)canbeassociatedwithAFterminationbyCA[10].In otherstudies[8],ithasbeenarguedthatthehighertheamplitude 1746-8094/$–seefrontmatter© 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.bspc.2013.02.002
Pleasecitethisarticleinpressas:M.Meo,etal.,Catheterablationoutcomepredictioninpersistentatrialfibrillationusingweightedprincipal ofthefibrillatorywaves(f-waves)observedonthesurfaceECG,the
morelikelyproceduralsuccess.
Inparallel,anotherlineofresearchaimsatnoninvasivemeas- ures of AF spatio-temporal complexity, with the underlying assumptionsthattreatmentmodalitiesshouldbechosenandther- apyoutcomecouldbepredictedonthebasisofthesemeasures.In [11],anoninvasivemeasureofAForganizationisassessedbythe normalizedmeansquareerror(NMSE)valuesbetweentheatrial activity(AA)signalanditsrank-3approximationsdeterminedby principal component analysis(PCA) in lead V1 [11]. Thisargu- mentissupportedbythehypothesisofacorrelationbetweenAF organizationandthenumberandinteractionsofatrialwavefronts throughtheheartsubstrate.ThechoiceofV1 is justifiedbythe factthatitpresentsthemaximumatrial-to-ventricularamplitude ratioamongallECGleads[12].In[13],CAperformancewasshown toinfluenceAFspatio-temporalorganization,anditseffectcanbe quantifiedbyvariationsinNMSEvalues.
Nevertheless,suchparametersare affectedbyseveralshort- comings.Inthefirstplace,someclassicalECG-baseddescriptorsare manuallycomputed[8,10],sotheyaresubjecttooperators’subjec- tivityandthuspronetoerrors.Furthermore,asmostofthemare measuredinonlyoneECGlead,theydonotaccountforinforma- tionthatmaybeprovidedbyotherelectrodes.Indeed,ECGanalysis isnotalwaysstraightforward,andvisualinspectiondoesnotcap- tureAFfeaturesunderlyingthewholeensembleofleads;hence, thelimitations ofclassicalsingle-leadtechniques,whichdo not fullyexploitmultileadECGspatialdiversity.However,AFspatio- temporalcomplexity as definedin [11] hasnotbeenshown to correlatewithCAoutcome.
Ourinvestigationfocuses onthepotentialapplication ofthe spatio-temporalorganizationofAAmeasuredonthestandardECG bytheNMSEindexasatooltodiscriminatebetweensuccessfuland failingCAproceduresbeforeapplyingthetherapy.Contributions providedfromtheeightindependentECGleadsareexpressedin terms onNMSE betweensuccessivesegmentsof theactual AA signaland theirrank-1 approximations computedby weighted principalcomponentanalysis(WPCA),andtheyarefinallycom- binedin asingleparametercapableofpredictinglong-termCA outcome.Thankstothisdecomposition,thespatialvariabilityof thestandardECGistakenintoaccount,andthemostsignificant ECGleadsarealsoautomaticallyenhancedbyassigningdifferent weightstodatabasedontheirestimatedrelevance.
2. Methods
2.1. Characteristicsandacquisitionmodalitiesofthe persistent-AFdatabase
Twentypatients(19males,60±11years)withamedianper- sistentAFepisodeduration of4.5months(2–84)wereenrolled in thepresent study. Theyall underwent CAat theCardiology DepartmentofPrincessGraceHospitalinMonaco,performedwith theaidofPruckaCardiolabandBiosenseCARTOelectrophysiology measurementsystems.Theyallgavetheirinformedconsent.Sur- face12-leadECGrecordingswereacquiredatthebeginningofthe procedure,atasamplingrateof1kHz.Anexampleofthesignal recordedontheleadV1foroneofthepatientsisshowninFig.1.
CAwasaccomplishedaccordingtothesequentialstepwiseprotocol [14],whosemajoractionsconsistin1)circumferentialPVisolation, 2)fragmentedpotentials’ablation,and3)non-PVtriggers,roofline andmitralisthmuslinerightatrialablation.
ThemostrecentHRSExpertConsensusStatementguidelinesfor CAtrials[14]recommendthatimmediatelyafterCAperformance, thereisathree-month“blankingperiod”duringwhichanyfibrilla- toryepisodesarenotregardedassymptomsofAFrecurrence,butas
0 200 400 600 800 1000 1200 1400 1600 1800 2000
−0.6
−0.4
−0.2 0 0.2 0.4
Time (ms)
Voltage (mV)
f − waves
T T T
QRS QRS QRS
Fig.1.ExampleofECGrecordingduringAFanditscharacteristicwaves.Boxes highlightTQintervalswhichareconcatenatedtoformtheAAsignalYAAinEq.(1).
aphysiologicalreactionduringrecoveryfromCA.Afterthisblank- ingperiod,ifthepatientremainsfreeofarrhythmiarecurrences, proceduralAFterminationisconsideredeffectivelyaccomplished.
ProceduralsuccessisdefinedasfreedomfromECG/Holterdocu- mentedsustainedAFrecurrence(>30s)duringfollow-up,afterthe 3-monthblankingperiod.ImmediatelyafterperformingCA,AFcan beconvertedeitherdirectlytoSRortointermediatetachyarrhyth- mia,exclusivelybyablationorafter anelectricalcardioversion.
For itsclinical interest, a long-termcriterion is adopted in our investigationtodistinguishbetweensuccessfulandineffectiveCA procedures.Inourexperimentalframework,afteramedianfollow- upof9.5months,CAwassuccessfullyaccomplishedinnS=13outof nP=20patients(65%),whereasnF=7procedureswerenoteffective.
Afollow-upofmmonthswasavailableatthetimeofouranaly- sis,wheremrangedbetween4and19monthsdependingonthe patient.
Some patientsreceived a pharmacological treatment subse- quent toCA procedure, mainlyamiodarone (forsome patients, solatolandflecaine).Threepatientsunderwentasecondablation.
In thiscase, only ECGsignals related tothefirstprocedureare takenintoaccountinourstudy.Asopposed topreviousstudies [8,9],terminationofAFduringCAwasnotachievedinallpatients.
Nevertheless,thisisnotdetrimentaltoouranalysis,sinceAFter- minationbyCAisnotpredictiveoflong-termoutcome[15],which istheeventwithclinicalinterest.
2.2. ECGpreprocessingandatrialactivitysegmentation
A fourth-orderzero-phase Chebyshev type IIbandpassfilter with−3dBattenuationbandbetween0.5Hzand30Hzhasbeen appliedtostandardECGrecordingsofourdatabase,whoselength isabout1min.ThispreprocessingstageallowsAFcontentenhance- ment,whosedominantfrequencytypicallyrangesbetween3and 12Hz, aswell as removal of baseline wandering and highfre- quencynoisesuchasmyoelectricartifactsand50Hzpowerline interference.Automaticdetectionof ECGfiducial pointsis then accomplished, in order to segment TQ intervals. R wave time instantsaredetectedonleadV1usingthePan–Tompkins’algorithm [16].Then,Qwaveonsetisdefined40msbeforethesubsequentR wave,thisbeingthetypicaldurationofthiswaveinthesecondi- tions(anabnormalQwavedenotespresenceofinfarct).Finally, TwaveoffsetisidentifiedwithanimprovedversionofWoody’s methodandautomaticallycomputedaftervisualinspectionand selectionoftheleadexhibitingthemostvisibleTwaves(ingen- eral,V2andV3)[17].Suchintervalsarefinallymean-correctedand
Pleasecitethisarticleinpressas:M.Meo,etal.,Catheterablationoutcomepredictioninpersistentatrialfibrillationusingweightedprincipal concatenated,soobtainingthe(L×N)datamatrixYAArepresenting
AAcontentonly:
YAA=[yAA(1)···yAA(N)]∈RL×N (1) wherevectoryAA(t)=[y1(t),...,yL(t)]TrepresentsthemultileadAA signalatthesampleindext,Lstandsforthenumberofleadsused, andNthenumberofsamplesoftheAAsignaly!(n)foreachlead
!=1,2,...,L.
TheredundancyinECGleads[18]promptedustodiscardlead IIIandGoldberger’saugmentedleads(aVR,aVL,aVF),sinceleadsI andIIcanfullycharacterizeheartelectricalactivityonthefrontal plane.Finally,all precordial leadshavebeen introducedtoo,in order torecordthe electricpotentialchanges in theheart in a cross-sectionalplane.ThisyieldsatotalofL=8leads,thatis,I,II, V1–V6.
2.3. Atrialactivitycomplexity
Severalparameters describingdifferentaspectsof AFspatio- temporal organization have been proposed and analyzed in previousworks,accordingtovariousdefinitionsofthisconcept.The rationaleistoinvestigateevidencesofsomeunderlyingstructure inatrialactivityduringAF.Thewidevarietyofdifferentcriteriathat havebeenproposedintheliteraturemakesitdifficulttocompare andinterpretallsuchindices.
ThedegreeoforganizationofAFwavefrontspropagatinginside theatria hasbeentraditionallyexaminedonintracardiacrecor- dings. In [19] thelevel of spatial correlation betweenmultiple activationsequencesiscorrelatedwithAFpresence,andenables selectionofantiarrhythmicdrugtherapyforSRmaintenance.In [20],AFmorphologycharacterizationbasedonPCAandautomatic clusteringprovidesa quantitativetool for AFclassification.The studydescribedin[21]alsoproposesmoreadvancedtechniquesfor featureextractionandSVM-classificationtoperformthesametask.
Otherapproachesfocusontemporalregularityofatrialactivations andassess AFcomplexity accordingtothelevelof beat-to-beat variability[22].Morerecently,time-frequencyanalysishasbeen appliedtoendocavitarianrecordingsin[23].Despitetheireffec- tiveness,suchstrategiesareallquiteinvasiveanddonotprovidea prioripredictionsofCAoutcome.
Bycontrast,noninvasiverecordingscaneasilyprovidemeasures ofheartelectricalactivity.Somenonlinearmeasuresbasedonsam- pleentropy[24]computedonsurfaceECGhavealsobeenexploited topredictspontaneousparoxysmalAFtermination[25].In[26],it isalsostatedthatmoreorganizedAFpatternsasquantifiedbythis indexpredictAFterminationbyelectricalcardioversion.Themain drawbackoftheseindicesisthattheyarecomputedinonlyoneECG lead,thuspotentialinformationaboutAFcomplexityprovidedby theremainingleadsisnotexploited.
Recentattempts toexploitECGspatialproperties havebeen madein[27]bycombiningfrequencyandcomplexitymeasures, allowingthedistinctionbetweenpersistentandlong-standingAF.
Also,in[28]wavefrontpropagationmapsextractedonBSPMrecor- dingshavebeenusedforvisualclassificationofAFcomplexitytypes accordingtoKonings’criteria[1]onBSPMrecordings.Basedonthis approach,aquantitativemultileadanalysisiscarriedoutin[11].
Thisstudyunderlinesthat AAspatio-temporal organizationcan beeffectivelyrepresentedbythefirstfewprincipalcomponents (PCs)determinedbyPCA,retainingmostofthetotalvariance,by quantifyingthesimilaritybetweentheprincipalsubspacesofthe AAsignalalongconsecutivetimesegments.Forsakeofcomplete- ness,themathematicaldescriptionofthiscomplexitymeasureis summarizednext,asitconstitutesanimportantingredientofthe presentwork.
ThemultileadAA signalYAA issplitinto Sequal-lengthseg- ments,eachcontainingNS=[N/S]samples,sothatYAA=[Y(1),Y(2),
..., Y(S)], with Y(s)=[y((s−1)NS+1), y((s−1)NS+2), ..., y(sNS)], s=1,...,S.TheAAsignalY(r) isexaminedinacertainreference segment r=/ s and decomposedby PCAaccordingtothelinear modelY(r)=M(r)X(r),r=1in[11].Subsequently,afixednumbern ofcolumnsM(r)n isextractedfromthemixingmatricescomputed byPCAinthisreferenceinterval.Suchcolumns,theso-calledprin- cipaldirections,weighttherelativespatialcontributionofthePCs totheECGleads.Afterthesesteps,theAAsignalisestimatedinall othersegmentss=/ rbyprojectingY(s)onthesubspacespanned bythecolumnsofM(r)n ,thusyielding:
Y(s,r)n =M(r)n [M(r)n TM(r)n ]−1M(r)n TY(s) (2) thatis,theorthogonalprojectionofY(s)onthespanofM(r)n .Hence, theapproximationqualitycanbeevaluatedbymeansofthenor- malizedmean square error NMSE(s,r)!,n between theinput signal y(s)! (t)onthe!thleadanditsprojection ˆy(s,r)!,n(t)foundinthe!th rowofEq.(2):
NMSE(s,r)!,n =
!
Nt=1[y!(s)(t)−yˆ(s,r)!,n(t)]2
!
Nt=1[y(s)! (t)]2 (3)
with!=1,···,L.In[11],themeanNMSEiscomputedbyassuming thefirstsegmentasareference(r=1),andaveragingEq.(3)over theremainingsegments(s=2,...,S).
Nonetheless,in[11]theNMSEintroducedinEq.(3)ismerely computedonasinglelead,sospatialvariabilitytypicalofmulti- leadrecordingsisnotentirelyexploited.Inparticular,despitethe proximityofleadV1totherightatrialfreewall,thereistheriskof notconsideringfurtherusefulinformationprovidedbyotherECG leads.Inaddition,asallaforementionedparameters,thissingle- leadindexprovedunabletopredictCAoutcome[13].
In order to render a more general perspective of AF com- plexityoverallleadsand providetheNMSEindexwithfurther clinicalvaluewithrespecttoCAoutcomeprediction,severalalter- nativestrategies havebeenputforward.In[29]somestatistical descriptorscombiningNMSEcontributionsfromseveralECGleads depending onAA signal variance were ableto assess thelevel ofspatio-temporalrepetitivenessoftheAAsignalduringseveral steps ofthe ablationand predictits outcome. NMSE computa- tionwasrepeatedforeachsegment,forallpossiblecombinations between estimatedand referencesegments r, s=1, ..., S, with r=/ s.Theanalysispresentedin[30]putsforwardthecomputa- tionofreduced-rankrepresentationsofAFcomplexitymeasures bymeansofthenonnegativematrixfactorization(NMF).Despite theiradvantages, thesestrategies proveeffectiveonly inshort- termprediction,sotheyarenotabletorenderthemechanisms ofelectricalremodelingoftheheartsubstratealteredbyCAover longerfollow-upperiodsanddiscriminatebetweentheclassesof interest.Resultsfromtheseworksencouragedustodesignamore robustmethodologycapableofselectivelyemphasizingthemost relevantcontributionsfromECGobservationstoimproveclassifica- tionaccuracyintheCAoutcomepredictioncontext.Moreprecisely, weaimatcharacterizingtheNMSEindexinamultileadframework soastopredictCAoutcomebyapplyingtheweightedPCAofthe atrialsignalmatrix.
2.4. Weightedprincipalcomponentanalysis
Asmentionedintheprevioussection,apossiblestrategyaim- ingattheexploitationofthemultivariatepropertiesofthestandard ECGconsistsinsearchingforareducedsetofuncorrelatedcompo- nentsretainingasmuchofitsspatialvariabilityaspossible.Inorder toachievethisobjective,PCAhasbeenwidelyappliedtotheECG, duetoitsnon-parametricnature,simplicityofimplementationand
Pleasecitethisarticleinpressas:M.Meo,etal.,Catheterablationoutcomepredictioninpersistentatrialfibrillationusingweightedprincipal versatility[31–34].However,PCAissometimesnotrecommended
inECGprocessing.Asitgivesthesamerelevancetoallobservations, thelow-rankrepresentationtendstobequitesensitivetooutliers andcanbecomeunstable.Thisissueaffectseverydecomposition methodbasedontheminimizationofacriterionfunctioninthe ordinaryleastsquares(OLS)sense.Toavoidtheselimitations,we putforwardamorerobustmultivariateanalysisaimingatdecom- posingthemultileadAAsignaldefinedinEq.(1)andsplitintoS segments.
Theapproachproposedistheweightedprincipalcomponentanal- ysis(WPCA),fittingthedatamodelY(r)byminimizingaweighted leastsquares(WLS)lossfunction[35]asintheschemedescribed next.AccordingtotheWLSapproach,eachentryoftheinputmatrix Y(r)definedinareferencesegmentrisseparatelyweightedwith afixed,nonnegativequantity.Theseleveragingfactorscanbecol- lectedina matrixW(r) havingthesamedimensionsasY(r).The generalformoftheWLSlossfunctioncanbewrittenas:
h(Y(r)|Y(r),W(r))= %(Y(r)−Y(r))∗W(r)%2F
=
"
L!=1
"
Nm=1
[w!,m(r) (y(r)!,m−yˆ(r)!,m)]2 (4)
where*denotestheHadamard(orelementwise)product,whereas theoperator||·||FstandsfortheFrobeniusnorm.Inourexperi- mentalframework,theoutputmodelY(r)=M(r)X(r)consistsofthe WPCAvectorscontainedintheL×nmatrixM(r),andthen×NS matrix X(r) representing thePCs, stored in decreasing order of variance.AsinclassicalPCA,someorthogonalityconstraintsare imposedonM(r)andX(r)inordertoreducetheambiguitiesinthe model.
2.5. Assignmentoftheweightmatrix
Specialconsideration mustbepaidtotheassignmentofthe weightscollectedinmatrixW(r).Indeed,anaccuratechoiceofthese valuescangiverisetoakindoffilteringactionwhichenhances not only theleads, but also thetime samplesgiving themost content-bearingcontributionswhilediscardingthosethatdonot yieldsignificantinformationorcanpolluteatrialobservations.In ourapplication,alltemporalsamplesonthesameleadareequally treated.Also,weaimatemphasizingleadswithmorestableand regularwaveformswhilereducingtheinfluenceofthosecharacter- izedbyhighertemporaldispersion,quantifiedintermsofenergy.
WeassumethattheinputsignalY(r)canbemodeledas:
Y(r)=Y(r)S +Y(r)N (5)
namely,asthesumofa meaningfulcomponentY(r)S (describing AFinourapplication)andanoisycomponentY(r)N,duenotonly todataacquisitionnoise, butalsotoelementsdiscardedbythe reduced-rankWPCA-approximation.OurhypothesisisthatY(r)N = (Y(r)−Y(r)S )ischaracterizedbyahighdegreeofspatio-temporal variabilitywhichcanalter orhideinformativeelementscoming fromYSineachlead.Thistermcanberegardedastheargumentof theWLScriteriondefinedinEq.(4)tobeminimizedaccordingto thealgorithmdescribedinthesequel.Hence,onewayofreducing itsinfluenceontheoverallsignalisbyweighingeachleadbythe inverseofAAsignalvariance,thusgivingmoreimportancetothe leastpowerfulelectrodes.Asaresult,eachrowofW(r)isweighted bytheinverseofthestandarddeviation"(r)! associatedwiththe correspondinglead!=1,...,LinY(r)andcomputedoneachsegment r=1,...,S:
W(r)=[("1(r))−1("2(r))−1...("L(r))−1]T1 (6)
where1isarowvectorwithTunitentries.Itiswellworthnoting thatclassicalPCAisaspecialcaseofWPCA,wheretheelementsof theweightmatrixareallequalto1.Onceaweightmatrixhasbeen chosen,WPCAcanbecarriedoutusingthealgorithmsummarized intheAppendix.Hence,thechoiceofdecomposingeachreference timeintervalinkeepingwiththeWPCAmodelY(r)=M(r)X(r).Inour application,thismodeliscomputedineachreferencetimeinter- val.Insuchacontext,AAsignalestimationqualitypersegmentis quantifiedbytheNMSEdefinedinEq.(3).Differentweightmatrices W(r)willgenerallyleadtodifferentmodelsM(r)X(r),thusresulting indifferentNMSEvalues.Thepredictivevalueofdifferentformsof W(r)willbetestedinSection3.
2.6. AssessingatrialactivitycomplexityfromtheNMSEvalues AfterWPCAperformance,weinvestigatehowtoproperlycom- bine NMSE values computed on each ECG lead and condense informationinauniquepredictorofCAoutcome.Withreferenceto Section2.3,AFcomplexityevaluationineachsegmentisfollowed bythecomputationofthemeanvalue#!,nandthestandarddevi- ation"!,n oftheNMSEinEq. (3)overallpossiblecombinations ofestimatedandreferencesegments(s,r), foreachlead![29].
Parameter#!,nassessesglobalsegmentestimationperformance, whereas"!,nquantifiesAForganizationinter-segmentvariability.
Finally,contributionsfromallleadsanalyzedarecombinedintothe interleadNMSEweightedsum:
#
#n=
"
L!=1
#!,n
"!,n2 /
"
L!=1
1
"!,n2 (7)
whoseweightsarerepresentedbyNMSEinversevariancevalues 1/"!,n2 perlead; contributionscomingfromECGleadsrendering more regularand less dispersive patterns are consideredtobe morerelevant.Afurtherinterpretationof"!,ncanbegiveninterms ofuncertainty:lowvaluesofthisparameterdepictamorestable reconstructionacrosstimesegments,whereashighvaluesdenote higherprojectionerroruncertainty.Accordingly,leadsguarantee- inga morerobust AA contentcharacterization have a stronger influenceinthecomputationoftheoutputdescriptor.Thechoiceof suchweightscanbefurtherjustifiedifweconsiderthatthecom- plexityinformationisreflectedontheensembleofECGleadsas asetofindependentrandomvariables.Thebestlinearminimum- varianceunbiasedestimatorofthecomplexitydescriptorwillthus begivenbytheweightedmeanofEq.(7)[36].Asaresult,greater weightisgiventovaluescomingfromlower-variancedistributions.
Theflowchartresumingthemainprocessingstagesofourmethod isrepresentedinFig.2.
2.7. ChoiceofNMSEcharacteristicparameters
SomeconsiderationsaboutthetuningofNMSEcharacteristic parametersshouldbementioned,i.e.,thenumberSofAAsegments whichneedtobeprocessed,besidesthenumberofspatialtopogra- phiesnusedforAAsignalestimation.ConcerningthevalueofS, experimentalevidencein[29]showsthatinmostpatientsNMSE decreasesandremainsconstantafteracertainthresholdSvalue.
SincetheerrorvariationflattensfromS≈4segments,wesetthis valuepriortosignaldecomposition.Thischoiceisalsosupported inthepresentstudybytheevolutionof#
#
WPCA8 asafunctionof thenumberofsegmentsS,assumingthattheirsizeisfixedand equaltoNSandtakingintoaccountconstraintsderivedfromthe ECGrecordinglengthavailableinourdatabase.AsFig.3shows,the indexkeepsquiteaconstantvaluewhenSincreases.Inaddition, evenwhensegmentlengthNS changes,parametervariationsare quitelimited(below10%),asshowninFig.4.Thisresultconfirms therobustnessofthepredictorputforwardtothechoiceoftuningPleasecitethisarticleinpressas:M.Meo,etal.,Catheterablationoutcomepredictioninpersistentatrialfibrillationusingweightedprincipal Fig.2.FlowchartofthealgorithmyieldingtheproposedCAoutcomepredictor
#
#WPCA8.Fig.3. Evolutionof
#
#WPCA8asafunctionofthenumberofsegmentsS.parametersandsupportsourchoiceofsettingauniquevalueforS forallpatientsinordertosimplifythealgorithm.
Concerning estimation performance, as WPCA rationale assumesthatthefirstPCretainsthemostoftheAAglobalvariance, therank-1approximationofAAobservationsbyprojectingthem overthedominantPCiscomputed.Indeed,suchareconstruction allowsnotonlyunderliningthemostdescriptivecomponentsin terms of variance, but also suppressingirrelevant and/ornoisy
Table1
AssessmentofWPCAconvergencecharacteristics:averagenumberofiterationsfor convergence$=10−5.
NS s
1 2 3 4
1000 96 84 114 93
2000 92 100 92 96
3000 84 90 92 103
4000 85 98 101 112
Fig.4. Evolutionof
#
#WPCA8asafunctionofthenumberofsamplespersegmentNS.elementsthatcandeterioratesignalcontent.Furtherexperimental resultsarepresented later inthepaper(Section3)and support thechoiceofsettingn=1,sothecorrespondingsubscriptwillbe omittedinthesequelforconvenience.Finally,asWPCAisaniter- ativealgorithm(describedintheAppendix),astoppingcriterion hasbeenintroduced,basedona convergencetolerance$=10−5 defined beforeEq. (9). In order toassess WPCA computational Table2
Assessmentof WPCAconvergence characteristics:averagefinalvalue ofWLS convergencecriterionC*(Eq.(9)afterconvergence,with$=10−5)[n.u.].Values normalizedbyascalingfactorequalto10−6.
NS s
1 2 3 4
1000 9.0 8.8 9.1 8.9
2000 8.8 9.0 8.6 9.2
3000 8.6 9.1 9.1 8.7
4000 9.0 8.9 8.9 9.0
Pleasecitethisarticleinpressas:M.Meo,etal.,Catheterablationoutcomepredictioninpersistentatrialfibrillationusingweightedprincipal Table3
InterpatientstatisticalanalysisandCAoutcomepredictionperformance(n.u.:normalizedunits).
SuccessfulCA FailingCA p-Value AUC Bestcut-off
#
#WPCA8[n.u.] 70.76±17.74 37.54±20.01 0.0013 0.91 40.64#
#WPCA12[n.u.] 63.03±18.12 61.64±20.87 0.88 0.47 53.76#
#PCA8[n.u.] 65.68±19.27 37.59±21.88 0.0082 0.84 45.57#
#PCA12[n.u.] 65.85±28.41 44.09±27.49 0.12 0.76 45.57NMSEWPCA8[n.u.] 54.24±25.89 50.24±23.04 0.74 0.57 45.64
NMSEWPCA12[n.u.] 85.67±14.62 86.78±17.41 0.75 0.54 95.32
NMSEPCA8[n.u.] 45.56±26.93 30.35±26.93 0.19 0.64 49.03
NMSEPCA12[n.u.] 67.80±22.11 80.73±14.48 0.18 0.69 69.00
D(V1)[mV] 0.08±0.03 0.06±0.01 0.03 0.80 0.05
SampEn(Ls,rs(A))[n.u.] 2.82±0.39 3.06±0.43 0.21 0.70 3.12
SampEn(Ls,rs(B))[n.u.] 2.42±0.38 2.66±0.43 0.20 0.70 2.73
SampEn(Ls,rs(C))[n.u.] 2.14±0.37 2.39±0.42 0.20 0.70 2.44
AFCLV1[ms] 139.63±19.66 121.75±23.83 0.09 0.71 129.87
load,thenumberofiterationsforconvergenceandthefinalvalue of WLS convergence criterion C* (introduced in Eq. (9) in the Appendix)arecomputedforafixedNSvalue.Thesevaluesarefirst computedoneachsegments=1,...,4andthenaveragedoverthe 20-patientdatabase.TestresultsarereportedinTables1and2.
Wecanremarkthat bothparametervalues donot significantly changewhenpassingfromasegmenttothefollowingone,and thatcomputationalburdenisrelativelylowintermsofiterations.
Asimilar resultis obtainedwhen considering variationsin the numberofsamplespersegmentNS.Thisfurtherevidencesupports therobustnessofourmethodtothechoiceoftuningparameters, asthealgorithmconvergestosatisfactoryresultsinallcases.
2.8. Statisticalanalysisandclassificationperformanceassessment As displayed in Table 3, categories under examination are referredtoas“SuccessfulCA”and“FailingCA”.Allparametersare expressedasmean±standarddeviation.First,Lilliefors’testwas runtoverifydatanormality.Differencesbetweensuccessfuland failingCAprocedureswerestatisticallydeterminedbyanunpaired Student’st-testifdataweresampledfromaGaussiandistribution, aWilcoxonranksumtestotherwise.Thep-valuesoutputbyeach unpairedtestareobtainedunderaconfidencelevel˛=0.05,and theyarereportedinTable3aswell.Binaryclassificationaccuracy ofeachfeatureisquantifiedbytheareaunderitsreceiveroperator curve(ROC),orareaundercurve(AUC),whosevalueiscorrelated withthemaximizationofsensitivityandspecificity,i.e.,thetrue positiveandtruenegativerates,respectively.
3. Results
Our8-leaddescriptor#
#
WPCA8iscomparedwithits12-leadcoun- terpart##
WPCA12.Moreover,thefinalweightedmeanofNMSEvalues hasalsobeencomputedforeachleadsubsetafterperforminga rank-1approximationbyclassicalPCA,thusobtaining#
#PCA8 and#
#PCA12,respectively.Acomparisonbetweenmultileaddescriptors andconventionalsingle-leadmethodsisdrawnaswell.Accord- ingly,AAamplitudeD(V1)iscomputedonleadV1accordingtothe algorithmproposedin [37,38].Moreover,atrialfibrillationcycle length (AFCL), widely known as a predictor of AFtermination byCA,isalsodeterminedonthesameelectrode. Itsmeasureis manuallydeterminedas describedin [39].Morespecifically,its valueisobtainedbyaveragingtemporaldistancebetween30con- secutivef-waves, thusgiving AFCLV1 as output.In addition,we examinea single-lead complexity measure basedonthe NMSE valuecomputedonV1eitherbyapplyingWPCAorclassicalPCA.
Suchdecompositionsareaccomplishedboth onthefullensem- bleofECGleadsandthereduced8-leadsubset,thusresultingin NMSEWPCA8,NMSEWPCA12,NMSEPCA8 andNMSEPCA12,respectively,
asoutputs.Aparallelwithanon-linearcomplexitydescriptor,the sampleentropy SampEn [24,26,40],hasbeendrawnas wellon V1.Twoparametershavetobetunedpriortoitscomputation:Ls
andrs.ParameterLsisdefinedasthelengthofthesequencesthe ECGrecordingissplitinto.Suchsegmentsarethencompared,and thetoleranceforacceptingmatchesisassessedbythethreshold rs.Thisparameterischosenasa fractionoftheAAinputsignal standarddeviationonV1,denoted"V1,soastoassurethetransla- tionandscaleinvarianceofSampEn.Parametervalueshavebeen tunedaccordingtotheguidelinesgivenin[24],sowesetLs=2 besidesthreevaluesofrs,namely,rs(A)=0.1"V1,rs(B)=0.15"V1and rs(C)=0.2"V1.
Thegeneralization powerof ouranalysis toanindependent datasetisvalidatedbymeansofaleave-one-outcross-validation technique.More precisely, AUCvalues have beencomputed on everypossiblesubsetof19patients,andthenaveragedoverthe 20subsets.AUCvaluesdescribingtheclassificationpowerofeach descriptorareshowninTable3,besidesthecorrespondingoptimal cut-offpoints,providingboththehighestsensitivityandthehigh- estspecificityonthewhole20-patientdatabase.Fig.5plotsthe
2 3 4 5 6 7 8
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
L
AUC
PCAWPCA
Fig.5.AUCvaluescharacterizing
#
#WPCALpredictionperformanceasafunctionof thesizeLofthesubsetofthe8independentECGleads(S=4,n=1).WPCA:rank-1 decompositionoftheatrialsignalintheECGleadsubsetsaccordingtotheWPCA approach;PCA:rank-1decompositionoftheatrialsignalintheECGleadsubsets accordingtothePCAapproach.Pleasecitethisarticleinpressas:M.Meo,etal.,Catheterablationoutcomepredictioninpersistentatrialfibrillationusingweightedprincipal Table4
ECGleadsubsetswithoptimalpredictionperformanceof
#
#WPCA8.Numberofleads(L) Leads
2 I,V1
3 I,II,V2
4 I,V2,V3,V5
5 [I,II,V2,V4,V5]
[I,II,V1,V3,V6]
6 I,II,V2,V3,V4,V6
7 I,II,V1,V2,V3,V4,V5
AUCvaluesdescribingtheclassificationperformanceof#
#
WPCALas afunctionofthenumberLofleadsretainedintheanalysis.For eachvalueofLrangingfrom2upto8,#
#PCALvaluehasbeencom- putedforall8!/((8−L)!L!)possibleensemblesofleads.Foreach leadcombination,CAoutcomepredictionperformancehasbeen assessedfromthecorrespondingvaluesof##
WPCAL,andvalidated bytheleave-one-outtechnique.ForeachsizeL,theminimum,max- imumandmeanAUCvaluesoverallL-leadsubsetswereobtained asafunctionofthesubsetdimensionL,andtheirrelatedranges ofvaluesaredisplayedinFig.5.Theleadcombinationswiththe bestpredictionperformanceforeachsubsetdimensionareshown inTable4.TheapplicationofPCAonasinglelead(L=1)hasbeen excludedfromthistest,sinceinthiscasethemethodisequivalent tosingle-leadanalysis.In ourapplication,wedeal withmultivariatedecomposition techniquesbasedonthemaximizationofthevarianceoftheAA signal,conveyinginformationaboutAFspatio-temporaldistribu- tion.Hence,anothercrucialpointofourinvestigationistheeffect ofsuchtechniquesonAAsignalenergycontent.Morespecifically, theinputAAsignalvariancehasbeendeterminedoneachECGlead, yieldingtheatrialpowerdistributionrepresentedbythevector
!2AA=["AA21,"AA22,...,"AA2L]T,withL=8.Effectsofthemultilead weightingschemeonthedecompositionsoftheobservedAAsig- nalare alsocompared to those obtainedby standard PCA.The rank-1approximationsYcomputedbyWPCAandPCAyieldatrial powerdistributionvectors!2WPCAand !2PCA,respectively.These
I II V1 V2 V3 V4 V5 V6
0 0.5 1 1.5 2 2.5 x 10−3
Lead Signalvarianceσ2(mV2)
σ2AA σ2PCA σ2WPCA
Fig.6.EffectsofthemultileadweightingschemeonAAreconstruction.!2AA:vari- anceoftheinputAAsignalperlead;!2PCA:varianceperleadoftherank-1AAsignal approximationbyPCA;!2WPCA:varianceperleadoftherank-1AAsignalapproxi- mationbyWPCA.
I II V1 V2 V3 V4 V5 V6
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Lead
AUC
σAA2 σPCA2 σWPCA2
Fig.7.AssessmentofCAoutcomepredictionperformanceofsingle-leadenergy descriptors.!2AA:energyoftheinputAAsignalperlead;!2PCA:energyperleadofthe rank-1AAsignalapproximationbyPCA;!2WPCA:energyperleadoftherank-1AA signalapproximationbyWPCA.
parameters have beencomputed over thewhole persistent AF databaseandaveragedoverallpatients;theirspatialdistribution is plotted in Fig.6. Thisfigureevaluates theenergy contentof data reconstructions computedby each decomposition, as well astheircapability ofeffectively approximatingtheoriginalsig- nal.Followingthisline,thisevaluationhasbeenaccomplishedin theframeworkofCAoutcomepredictionaswell.Inparticular,we testedwhetherAAsignalenergy"AA2 associatedwitheachleadcan effectivelyperformasapredictoroftheablationresult;hence,the quantificationoftheirclassificationaccuracyoneachECGleadby meansoftheAUCcriterion,whosevaluesaredisplayedinFig.7.
Thesameanalysisis ledontheenergyvaluescomputedonthe rank-1approximationsoutputbyPCAandWPCA,"PCA2 and"WPCA2 , respectively,alsoplottedinFig.7.Finally,furthertestsconfirmthe validityofthemodelintroducedinEq.(5)byassessingCAoutcome predictionperformanceonthebasisofdifferentdefinitionsofthe weightmatrixW(r)aswillbediscussedinSection4.5.
4. Discussion
This work investigates noninvasive measures of AA spatio- temporal variabilityand theirlinktoCAoutcome predictionin persistent AF. The main resultscan be summarized as follows.
Firstly,spatialvariabilityofthestandardECGprovestobeause- fultooltodescribeAFcontentandofferawiderperspectiveabout theevolutionofthediseaseduringCA,thushelpingitsoutcome prediction.Inthesecondplace,anindexconventionallyemployed asaclassifierofAForganizationtypeishereincharacterizedsoas toassessCAeffect.InformationaboutAAsignalcomingfrommul- tipleECGleadsisexploitedbyapplyingWPCA,whichefficiently compressesAFcontentinaunique,significantPC,thereby min- imizingtheimpactofpollutingsignalcomponents.Theweights usedinWPCAseemtoautomaticallyenhancetheroleofthemost descriptiveECGleads,unlikePCA,whichequallyweightsallECG leads.Thisfilteringactionisalsoperformedbyselectingthe8-lead ensembleintroducedinSection2.2,sosuppressinglineardepend- enciesbetweencertainleadsduetotheirspatiallocation.These aspectsarediscussedinmoredetailinthesequel.
Pleasecitethisarticleinpressas:M.Meo,etal.,Catheterablationoutcomepredictioninpersistentatrialfibrillationusingweightedprincipal 4.1. CAoutcomepredictionintheWPCAmultileadframework
Our experimental results point out the advantages of the multileadstrategywhichconsiderablyoutperformsconventional predictorscomputedinonlyoneECGlead.Concerningsingle-lead AFcomplexitymeasuresdeterminedbyPCAondifferentsetsofECG leads,i.e.,NMSEPCA8 andNMSEPCA12,notonlystatisticallysignifi- cantinterpatientdifferencescannotbeobserved,butAUCvalues relatedtotheirdiscriminationcapabilityarealsoextremelylow.
Similarconclusionscanbedrawnwhenexaminingtheequivalent parametersobtainedinV1whenapproximatingdatabymeansof WPCA(NMSEWPCA8,NMSEWPCA12).Theseresultscouldbeexplained bythelimitedoutlookofsingle-leadcomplexitymeasures,which ignoreinterleadrelationships. Relevantinformation fromother electrodesisneglected,thusreducingdiscriminationcapabilities.
Furthermore,asleadV1isclosetotherightatrialfreewall,there istheriskofneglectingusefulinformationaboutotherimportant anatomicalareas,suchastheleftatrium(LA)andthePVs,which playacrucialroleinAFinitiationandmaintenance[41].In[42], itisshownthatV1istheleadthatbestexplainsleftatrialactiv- ityintwosubjectsaffectedbyatrialtachycardiaconfinedtothe LA.However,AFmechanismsaregenerallymorecomplex,andour resultsinSection4.5indicateinanycasethat,concerningCAout- comeprediction,leadV1 doesnotdepictimportantinformation aboutablationeffectsthatcouldbepresentinotherleads.Also,con- cerningnonlinearAFcomplexityindicessuchassampleentropy, nostatisticallysignificantinterclassdifferencescanberemarked, regardlessofthevalues of tuningparameters.Notonly sample entropyindexisaffectedbythesameshortcomingstypicalofthe othersingle-leadfeatures,butitisalsonecessarytosetvaluesof itstuningparameterspriortoitscomputation.
By contrast,by meansof WPCA, ECG spatialdiversity high- lightsstatisticallysignificantdifferencesbetweenthecategories examined. As displayed in Table 3, higher values of the mul- tilead descriptor #
#
WPCA8 are significantly correlated with CA success.AsitsmathematicaldefinitioninEq.(7)shows,theNMSE inter-segmentvarianceprovidesa quantitativecriterion forthe assessmentofthespatio-temporalvariabilityoftheAAsignal:leads exhibitingmorestableandrepetitivepatternsgiveamorerelevant contributiontotheweightedmean##
WPCA8,sotheyhavestronger influence.Theinter-segmentvariance actsasa leadselector: it enhancesECGelectrodeswherenotonlythesignalshowsthemost stablewaveform,butitisalsolikelytoyieldamoreaccuratesignal estimation,withalowerdegreeofuncertainty.4.2. AcomparisonwithstandardclinicalpredictorsofCAoutcome Inthefirstplace,selectionofpatientstobetreatedbyCAis guidedbyconsiderationsaboutsomeclinicaldata,suchasLAdiam- eterandAFduration.Indeed,itiswidelyknownthatwhentheLA ismarkedlydilatedCAislesslikelytobeeffective,asalargerLAis linkedtoamoreadvanceddegreeofthepathology.Similarremarks canbemadeaboutAFduration,correlatedwithitschronification level[14].In[43]itisdemonstratedthatCAoutcomepredictionin paroxysmalAFisnotablyimprovedbytheknowledgeaboutsome features,primarilythepresenceofnon-PVdriversanddominant frequenciesbothinrightandleftatrium.Nevertheless,themost oftheseparametersareinvasivelyacquiredandknownonlyatthe momentoftheprocedure.Thismotivatestheinterestinpredic- tivefeaturesthatcanbeextractedwithoutrisktothepatientina cost-efficientmanner,asthosederivedfromtheECG.
AnalysisoftheAAamplitudeasrenderedbyD(V1)arousesdif- ferentremarks.Actually,wecannoticethesatisfactoryabilityto distinguishbetweensuccessful and failingablations, as wellas theeffective reproduction ofresults manuallyreported in pre- viousworksusingadifferentpersistentAFdatabase[8].Such a
descriptorcaneffectivelycaptureAAsignalamplitudecharacter- isticsifthepatternissufficientlyregularand f-wavesareeasily detectable,althoughinterpolationoperationscanbehamperedby residualspuriouspeaksortooirregularpatterns.Whatismore,no informationaboutAFspatio-temporal repetitivenessisprovided bythisfeature.
The study presented in [39] assesses the predictive role of AFCL measured on surface ECG for CA of persistent AF.
However, as its value is manually acquired, there is a lack of reproducibility and prediction reliability. Moreover, experi- mental results show that such parameter, defined as AFCLV1, does not underline statistically significant differences between the categories under examination using lead V1. Further- more, as it is usually determined in only one electrode, it is affected by the shortcomings typical of single-lead predic- tors.Inaddition,correlationbetweenECG-basedparametersand intracardial measures hasnot beenconfirmed in somestudies [44].
4.3. WeightedandstandardPCA:acomparison
EventhoughstandardPCAiscapableofdiscriminatingbetween successfulandfailingCAproceduresaswell,resultsconcerning
#
#WPCA8showthatclassificationqualitycanbefurtherimprovedby ouraprioriknowledgeaboutatrialobservationsintheformofthe weightsusedinWPCA.Onfurtheranalysis,AAstandarddeviation measuredoneachECGelectrodeprovestobeareliableindex,since itdoesnotonlyweightAFtemporaldispersion,butitisalsoasta- tisticalmeasureofuncertainty.Indeed,ifAApatternsoncertain leadsare excessivelyirregular and/orvariable, thecorrespond- inginversestandarddeviationvaluesautomaticallyreducetheir influence.Thisselectiveaction seemstoboostthecompression powerofthedecomposition.Morespecifically,theeffectofpossi- bleredundanciesisalreadyreducedbeforecomputingtheiterative minimizationalgorithm byselectingthe8 linearlyindependent ECGleads,sothatthemostdiscriminantAAcomponentsareput intoevidencemoreeasily.InFig.5theAUCcriterionquantifiesthe classificationperformanceof#
#
WPCALand##
PCALasafunctionofthe numberofECGleadsexploitedforthepredictionselectedamong the8independentleads.ClassificationresultsobtainedusingWPCA outperformthosebyPCA,especiallyassizeLincreases.Thisfigure alsoconfirmsthebenefitsderivedfromthespatialvariabilityof thestandardECG.Thehigherthenumberofleadsemployed,the moreaccurateCAresultprediction,assessedbyhighermeanAUC values.Furtheradvantagesderivedfromtheweightingframeworkare displayedin Fig.6.First of all,it canbenoticedthat thetrend of!2WPCAvaluesisverysimilartothatof!2AA.Inaddition,!2WPCA values are closerto!2AA thanthose obtainedwhen performing classicalPCA(!2PCA),therebyquantifyingalowererrorofrecon- struction of theoriginal data. Energy values obtained after AA signalapproximation,eitherbyPCAorWPCA,arelowerthanthose computeddirectlyoninputdatabecauseofthelow-rankrepre- sentationeffect.ItcanbeinferredthatWPCAcanbetterpreserve energycontentoftheAAsignalandcondenseitmoreefficiently inasingle,maximum-variancePCthanconventionalPCA.Differ- encesbetweenthesedecompositionsintermsoftheamountof informationretainedbytherank-1approximationareparticularly evidentinV1andV2,whichrepresentthereferenceleadsforAF analysisinmedicalpractice,owingtotheirproximitytotheright atrium.NotehowWPCAsignificantlyenhances,inanautomated fashion,therelevanceoftheseleadsintheAAsignaldecomposi- tion.Theuseoftheseenergy-descriptorsinsingle-leadprediction doesnotprovidesatisfactoryresults,asshowninFig.7.Indeed, theirpredictionperformanceispoor,andalsohighlydependenton
Pleasecitethisarticleinpressas:M.Meo,etal.,Catheterablationoutcomepredictioninpersistentatrialfibrillationusingweightedprincipal theleadconsidered.Ingeneral,theseresultsconfirmtheneedfor
anadequatecombinationofatrialsignalcontributionsfromdiffer- entECGleadsinamorerobustmultileadframework,capableof filteringoutuninformativeAAsignalfeaturesandexploitingECG spatialvariability.However,theimportanceoftheircontribution issignificantlyreducedwhenapplyingPCA,thuslosingrelevant informationaboutAFenergycontentintheassociatedheartsites.
Conversely,theWPCAschemecaneffectivelyimproveCAoutcome predictionbyreinforcingthemostdiscriminativefeaturesofinput datathankstothepriorknowledgeaboutAAsignalenergydistri- bution.
4.4. Alternativedefinitionsoftheweightmatrix
The choiceof the inverse standard deviation values as W(r) weightsconfirmsourhypothesisaboutAA signalrepresentation illustratedinSection2.5.Conversely,otherweightingschemesare notabletogivecomparableclassificationresults.Forexample,the weightmatrixdependingonAAstandarddeviationvaluesperlead W(r)=["1(r),"2(r),...,"L(r)]T1doesnotmanagetoproperlyempha- sizeECGleadcontributions,thusshowingaweakpredictivepower (AUC=0.53,p-value=0.98).Theseresultsseemtocorroborateour AFmodel,as AAmaximum-power componentsseemunableto selectivelyenhancethemostinformativecontributionsbymeans oftheweightingstructure.
Infurtherexperiments,alternativeweightmatricesW(r) have alsobeen tested. For instance,one of the attempted strategies consists in giving more weight to leads better explained by a reduced-rankPCAapproximation,thusdefiningW(r) elementsas afunctionoftheinversevalueofstandarddeviationoftheerror between original data and rank-1 PCA approximation per lead (W(r)=[("1(r)
N)−1,("2(r)
N)−1,...,("L(r)
N)−1]T1).However,nosignif- icant differences between effective and failing CA procedures havebeenfoundinthiscase(AUC=0.67,p-value=0.69).Inother tests,wehypothesizethatW(r) componentsdependonthevalue ofthestandarddeviationitself(W(r)=["1(r)
N,"2(r)
N,...,"L(r) N]T1), although similar poor results are obtained (AUC=0.63, p- value=0.37). This leads us to conclude that focusing onnoisy componentsthatmaybepresentintheAAsignaldoesnotactu- allyimprovetheselectiveactionoftheweightingscheme,whereas consideringvarianceofthewholesignalgivesmoreemphasistoits mostinformativecomponents,thusimprovingpredictionaccuracy.
4.5. ECG-leadselection
Classificationperformanceof multileadCApredictors proves tobemoreaccuratewhenreduced-rankapproximationsarecom- putedonthesubsetof8independentleadsratherthanthewhole standard ECG.Thisresult is inlinewithrecent previousworks [38]. In clinical centers,all leadsof the standard ECG are ana- lyzedsothatprojectionsoftheresultantvectorsin2orthogonal planesatdifferentanglescanbecompared,soimprovingpattern recognition[18].However,WPCAcomputationoverthesubsetof 8independentleadsdefinedaboveseemstoincreasetheefficacy ofitsfilteringaction,aspartofredundantinformationisalready suppressedbeforethedecomposition,andthemodelcapturesthe naturalvariationunderlyingthedatamoreeasily.Theseconsid- erationsjustifythefactthatdescriptorsextractedfromthewhole ECG,namely,#
#
WPCA12and##
PCA12,areoutperformedbytheir8-lead counterparts.Table4displaysthegroupsofleadswhichbesthelpdiscrim- inatingbetweensuccessfuland failingprocedures,and givethe mostrelevantcontributiontothecomputation oftheweighted mean
#
#WPCA8 in the prediction scenario. Some considerations can be made about the role of certain leads in CA outcome1 2 3 4 5 6 7 8
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
n
AUC
Fig.8. AUCvaluesdescribing
#
#WPCA8predictionperformanceasafunctionofthe ranknoftheWPCAdecomposition.prediction. In particular, we can remarkthat leadV1 doesnot providethemaincontributiontoCAoutcomeprediction.Infact, otherleads,suchasI,II,V2,recurmorefrequently.Thisevidence isincontrastwithstandardmedicalpractice,anditcanbeprob- ably explained by theplacement of lead V1, not close enough tocriticalsitesresponsibleforAFgenesisandmaintenancesuch asthePVsandtheLA,commonlyacknowledgedaspotentialAF sources.Thisseemstobeconfirmedbytherecurrenceofatleast one lead closeto theleft side of theheart in each L-size sub- set,forinstance,V5 andV6.Wecanconcludethatthepresence ofleadsrepresentingheartelectricalactivityonmultipleplanes supportsthehypothesisthatclinicalinformationcomingfrommul- tipleelectrodelocationscanimproveablationoutcomeprediction ascomparedtoclassicalsingle-leadapproaches,asdiscussedin Section4.1.
4.6. Benefitsofreduced-rankWPCAapproximationstotheAA signal
Fig.8illustratestheadvantagesprovidedbydatacompression carried out by WPCA,and seemsto justify thechoiceof rank- 1approximations(n=1)madeinSection2.7.Indeed,AUCvalues relatedtoourmultileadpredictor#
#
WPCA8havebeencomputedby varyingthenumberofPCsnretainedintheWPCAtruncationin Eq.(2),whichrangesfrom1(thevaluesetforouralgorithm)to8 (full-rankdecomposition).ThequalityofCAoutcomeprediction considerablyworsenswhen increasingthetruncationrank,and#
#WPCA8 exhibitsweakdiscriminatingcapabilitieswhenassuming morethan4PCs.ThefewerPCsemployedinthedecomposition,the bettertheclassificationperformance,asifthedominantPCspre- servedthediscriminativepowerofthecomplexityindex.Indeed, noisyand/orredundantelementsaretypicallyascribedtothevery lastPCs,whilepreservingthemostrepresentativefeaturesofthe AAsignalinthedominantones.Asimilarbenefitofrankreduction wasobtainedin[38]inthecontextofshort-termCAoutcomepre- dictionbasedonamplitudeparameterscomputedfromlow-rank PCAapproximation.
4.7. LinkswithAFspatio-temporalcomplexity
Asourresearchdemonstrates, CAoutcome predictionbased onWCPAof multileadECGsignalsproves tobeeffective.How- ever,theproperlinkbetweentheNMSE-basedpredictorproposed andAFspatio-temporalcomplexitycannotbeestablishedinthe presentworkforlackofsimultaneousinvasiverecordings.Such a connectioncanonly besuggestedbytheresultspresentedin [11],whichshowedthecorrelationbetweentheNMSEmeasure in V1 and AF organization. Accordingly, the potential relation