"Non-invasive Prediction of Catheter Ablation Outcome in Persistent Atrial Fibrillation by Fibrillatory Wave Amplitude Computation in Multiple ECG Leads"

(1)

Availableonlineat

ScienceDirect

www.sciencedirect.com

CLINICAL RESEARCH

Non-invasive prediction of catheter ablation outcome in persistent atrial ﬁbrillation by ﬁbrillatory wave amplitude computation in multiple electrocardiogram leads

Prédiction non invasive du suivi au décours d’une ablation par cathéter d’une ﬁbrillation atriale persistante par l’analyse informatisée de l’amplitude des ondes de ﬁbrillation atriale sur un tracé ECG à multiples dérivations

Vicente Zarzoso

^a^,∗

, Decebal G. Latcu

^b

,

Antonio R. Hidalgo-Mu˜ noz

^a

, Marianna Meo

^c

, Olivier Meste

^a

, Irina Popescu

^b

, Nadir Saoudi

^b

aI3SLaboratory,UniversityofNiceSophiaAntipolis,CNRS,Sophia-Antipolis,France

bCardiologyDepartment,PrincessGraceHospital,Monaco

cElectrophysiologyandHeartModellingInstitute(IHULIRYC),Bordeaux,France

Received3November2015;receivedinrevisedform9January2016;accepted3March2016 Availableonline8July2016

KEYWORDS Atrialﬁbrillation;

Catheterablation;

Electrocardiography;

Fibrillatorywave amplitude;

Therapyoutcome prediction

Summary

Background.—Catheterablation(CA)ofpersistentatrialﬁbrillation(AF)ischallenging,and reportedresults arecapable ofimprovement. A betterpatient selection for theprocedure couldenhance itssuccessratewhileavoiding the risksassociated with ablation, especially forpatientswithlowoddsoffavorableoutcome.CAoutcomecanbepredictednon-invasively byatrialﬁbrillatorywave(f-wave)amplitude,butpreviousworksfocusedmostlyonmanual measuresinsingleelectrocardiogram(ECG)leadsonly.

Aim.—Toassessthelong-termpredictionabilityoff-waveamplitudewhencomputedinmul- tipleECGleads.

Abbreviations:AF,atrialﬁbrillation;AUC,areaunderthereceiveroperatingcharacteristiccurve;CA,catheterablation;CFAE,complex fractionatedatrialelectrogram;ECG,electrocardiogram;EGM,electrogram;LA,leftatrium;LR,logisticregression;PV,pulmonaryvein;

ROC,receiveroperatingcharacteristic.

∗Correspondingauthor.I3SLaboratory,UNS,CNRS,2000,routesdesLucioles,CS40121,06903SophiaAntipoliscedex,France.

E-mailaddress:zarzoso@i3s.unice.fr(V.Zarzoso).

http://dx.doi.org/10.1016/j.acvd.2016.03.002

(2)

Methods.—Sixty-twopatientswithpersistentAF(52men;meanage61.5±10.4years)referred for CA were enrolled. A standard 1-minute 12-lead ECG was acquired before the ablation procedurefor eachpatient.F-waveamplitudesindifferentECGleadswerecomputedby a non-invasivesignalprocessing algorithm,andcombinedinto amutivariateprediction model basedonlogisticregression.

Results.—Duringanaveragefollow-upof13.9±8.3months,47patientshadnoAFrecurrence afterablation.AleadselectionapproachrelyingontheWaldindexpointedtoI,V1,V2andV5 asthemostrelevantECGleadstopredictjointlyCAoutcomeusingf-waveamplitudes,reaching anareaunderthecurveof0.854,andimprovingonsingle-leadamplitude-basedpredictors.

Conclusion.—Analysing thef-waveamplitudeinseveralECGleadssimultaneouslycansigni- ﬁcantlyimproveCAlong-termoutcomepredictioninpersistentAFcomparedwithpredictors basedonsingle-leadmeasures.

MOTSCLÉS Fibrillationatriale; Ablationpar cathéter;

Électrocardiographie; Amplitudedesondes deﬁbrillation atriale;

Prédictiondusuivi

Résumé

Contexte.—L’ablationparcathéterdelafibrillationatriale(FA)persistantedemeureunprob- lème difficile et les résultatsrapportéssont perfectibles. L’affinement de la sélection des patientspourcetteprocédurepourraitaméliorerletauxdesuccèsenévitantlesrisquesliésà laprocédure,notammentchezlespatientsàfaibleprobabilitéderéussite.Lesuiviaudécours d’uneablationparcathéterpeutêtrepréditdefaçonnoninvasiveparl’analysedel’amplitude desondesdeFAmaisdestravauxprécédentssesontfocalisésprincipalementsurdesmesures manuellessuruneseuledérivationélectrocardiographique(ECG).

Objectifs.—Cetravailévaluelaprédictionàlongtermedusuccèsd’uneablationparcathéter d’uneFAparanalysenumériquedel’amplitudedesondesdeFAsurdesdérivationsECGmulti- ples.

Méthodeetrésultats.—Soixante-deuxpatientsprésentantuneFApersistanteontétéinclus, 52hommes,d’âgemoyen61,5±10,4ans.Pendantunsuivimoyende14±8mois,47patients n’ontpasprésentéderécidivedeFAaudécoursdecetteprocédure.UnECGstandardaété acquisavantlaprocédurechezchaquepatient.L’amplitude desondesdeFAsurdifférentes dérivationsECGaéténumériséeetunalgorithmenoninvasifutilisé,permettantd’établirun modèlemultivariéàpartird’uneanalyseparrégressionlogistique.Uneanalysebaséesurl’index deWaldsurlesdérivationsD1,V1,V2etV5permetdeprédireaumieuxlesuccèsd’uneablation parcathéteràpartirdel’amplitudedesondesdeFA,fournissantunesurfacesouslacourbe ROCà0,854,améliorantdonclaperformanced’uneprédictionbaséesurl’analysed’uneseule dérivationECG.

Conclusion.—L’analyse de l’amplitude des ondes de FAsur plusieurs dérivations ECG peut améliorerdefac¸onsigniﬁcativelaprédictiondusuccèsd’uneablationparcathéterchezdes patientssouffrantd’uneFApersistante.

Background

Atrialfibrillation(AF)isthemostcommonsustainedarrhyth- mia encountered in clinical practice [1]. Radiofrequency catheterablation(CA)ofpersistentAFisawell-established therapy,withproven efficacy in maintainingsinus rhythm duringfollow-up[2].Despiterecentsignificantprogress,CA for this form of AF yields less than perfect results, as it remainsacostly,time-consumingintervention,withriskof complicationsforthepatient.Hence,accurateselectionof long-termresponderstoCAiscrucialforimprovedpatient- tailoredmanagementofthiscardiaccondition.

The patient characteristics that correlate most to CA outcome are unclear [3—5]. The atrial ﬁbrillatory waves(f-waves)observedinthesurfaceelectrocardiogram (ECG) reﬂect the electrical behaviour of the atria in a non-invasivefashion,andtheiranalysisintime[6,7]orfre- quencydomains[8,9],orusingmoreelaboratecomplexity indices[10],hasbeenshowntocorrelatewithCAoutcome.

In previous studies, f-wave amplitude hasbeen measured manuallyinsingleleadsseparately(suchasII,aVForV1), thus neglectinginformationfromthe remainingleads that may berelevant for AF characterization.Indeed,the link betweenf-waveamplitudeandlong-termoutcomehasnot

(3)

beenclearlyestablished[7]orhasbeendemonstratedwith limitedaccuracyonly[6].

To overcome thesedrawbacks, thepresent study analysed whether the consideration of multiple ECG leads simultaneously could improve CAlong-term outcome pre- dictionbasedonautomatedf-waveamplitudemeasures.

Methods

Study population

Allpatientsunderwentradiofrequency ablationfor persistent andlong-standingpersistentAF [2]at PrincessGrace Hospital(Monaco).ThestudywasapprovedbytheInstitu- tionalCommitteeonHumanResearch,andallpatientsgave written informed consent. Anti-arrhythmic drugs (except amiodarone)werewithdrawnatleastﬁvehalf-livesbefore thestudy.Ratecontroldrugswereinterruptedjustbefore CA. Before the procedure, all patients had echocardio- graphicassessment oftheleftatrium(LA;anteroposterior diameterintheparasternallong-axisview;two-dimensional surfaceintheapicalfour-chamberview)andtheleftven- tricularejection fraction by Simpson’s biplanemethod. A computedtomography scanacquisitionof theLAwasalso performed foreach patientbeforetheprocedure. The LA three-dimensionalvolumewascalculatedandreconstructed onthecomputedtomographyscan.

Signal acquisition

For every patient, a 1-minute standard 12-lead ECG was recordedat asamplingrateof 977Hz immediatelybefore thestartofablation.ECGsignalswereacquiredonadigital electrophysiologicalrecordingsystem (PruckaEngineering, Inc.,Houston,TX, USA), including0.05 to40Hz bandpass and50Hznotchﬁlters.Forpatientswhounderwentrepeat procedures,onlytheECGrecordedbeforetheﬁrstinterven- tionwasconsideredinourpredictionanalysis.

Ablation procedure

TheproceduralapproachforLAarrhythmiaablationinthe study’s institution has been described elsewhere [11]. In short,weperformedcoronarysinuscatheterizationusinga decapolardiagnosticcatheter,doubletransseptalpuncture, systemicanticoagulation withheparin,withatarget acti- vated clotting time>350seconds, and electroanatomical mappingoftheLAwiththeCartosystem(Biosense-Webster Inc.,DiamondBar,CA,USA).Mappingandablationcatheters wereinserted transseptallyvia anon-steerable (Fast-Cath SL1; St.JudeMedical,Minnetonka,MN, USA)or steerable (Agilis,St.JudeMedical;or V-CasDeﬂect,Stereotaxis,St.

Louis,MO,USA)sheath.A20-polecircularmappingcatheter (Lasso2512;Biosense-WebsterInc.)wasusedtoassesspul- monaryvein(PV)potentials.

LAshellfor anatomicaldeﬁnitionwasdone witheither standard electroanatomical (Carto XP; Biosense-Webster Inc.)oradjustedfastanatomical(Carto3;Biosense-Webster Inc.)techniques.Image integrationwiththe LAcomputed tomographyscanreconstructionwasalwaysused.Detailed mappingofelectricalactivationoftheLA(and,inselected

cases, of the right atrium) was performed, with visual annotation of complex fractionated atrial electrograms (CFAEs) [12]. Ablations were carried out in a stepwise manner, with endpoints of circumferential PV disconnec- tion, ablation at CFAE sites and block across the lines (ifperformed). Inall procedures,the operator systemati- callydeliveredpoint-by-point ablation lesionsfor at least 30seconds to create a contiguous antral circumferential linearoundthePVpairs.Catheter—tissuecontactwasopti- mizedbeforeeachradiofrequencydelivery,usingcatheter motion on fluoroscopy, near-field electrogram (EGM) stability,impedancedrop duringradiofrequency deliveryand morphological EGM changes suggestive of lesion creation [11]. Whenever available, contact force was also taken intoaccounttooptimizecatheter—tissuecontact.Ifasig- nificant lesion (basedon localEGM modification) wasnot obtained,reablationatthesamesite,withfurtheroptimiza- tionof contact(at times requiring theuse of a steerable sheath) and energy increase, was performed. Irrigated radiofrequencywasdeliveredwithaStockert70generator (Biosense-Webster Inc.),a 42^◦Climiting temperature and 30—40Wfortheendocardialpartoftheline.Baselineirri- gationflowwas17to30mL/min,withanincreaseto60mL incaseofexcessiveheating.IfPVisolation(entryblock)was notobtainedattheendofthecircularlesion,thelineswere remappedandthegapsreablated.Ifneeded,furtherlesions guidedbythecircularcatheterweredelivered.

IncaseofongoingAFafterPVisolation,additionallesions targetingfractionated EGMs inthe LA aswell asthe roof and,inafewcases,theleftisthmuslines(withendpoints ofblock across the line) were performed. Further lesions weredelivered,inselectedcases,withinthecoronarysinus (20W) and,in some cases without AF termination,in the rightatrium,targetingfractionatedEGMs.

AF termination during ablation was defined as sinus rhythm resumption or its change to a stable atrial tach- yarrhythmia. Nevertheless, this was not a procedural endpoint, and operators ended the procedures after PV isolation,ablation of theannotatedCFAE sites and,when appropriate,blockacrossthelines.Associatedatrialtach- yarrhythmia ablation was performed, and the critical isthmusorfocaloriginwasspecificallytargeteduptosinus rhythmresumptionandreconfirmationofPVisolationbefore catheterwithdrawal.

IncaseswithoutAForatrialtachyarrhythmiatermination after catheterwithdrawal, a loading dose of amiodarone (30mg/kg) was administered, unless contraindicated. An electricalcardioversion(150to200J,repeateduptothree times under general anaesthesia) was performed if the arrhythmiawasongoing24to48hoursafterwards.

Follow-up

After the 3-month blanking periodrecommended by currentguidelines[2],patientswerefollowedforclinicaland asymptomatic recurrences. Follow-up wasperformed in a

‘‘real-life’’setting,byregularvisitstothetreatingcardiol- ogist,withrepeatedECGand24-hourHoltermonitoringin allcases(every3monthsduringtheﬁrstyearafterablation;

every6monthsafterwards).Supplementarydocumentation by ECG or Holter was sought in case of recurring symptoms suggestive of arrhythmia. Any recurring, sustained

(4)

(>30seconds),symptomaticAForﬂutterwasconsideredfor arepeatprocedure. Absenceof any AFrecurrenceduring follow-updeﬁnedtheCAsuccessgroupofourstudy,while patientswithdocumentedAFrecurrencesafterthelastpro- cedureconstitutedtheCAfailuregroup.

Signal-processing and statistical analysis

F-wave amplitude computation

Thesignal-processingalgorithmusedtocomputethef-wave amplitudein each lead is illustrated in Fig.1. First, ECG ﬁducialpointsweredetectedtoproperlysegmentTQinter- vals,whereatrialactivitycanbemeasuredfreefromQRST complexesofventricularinterference(Fig.1,top).R-wave timeinstantswerelocatedonleadV1byapplyingthePan- Tompkins algorithm [13], and Q-wave onsets were simply obtainedbysubtracting40ms,atypicalventricularactiva- tiontime.FromtheleadwherethemostprominentT-waves couldbevisuallyidentiﬁed,T-waveoffsetswereestimated byanadaptedWoody’smethod[14];then,thesegmented intervalsweremeancentredandconcatenated(Fig.1,middle).Intheconcatenated TQsegments, thelocalmaxima weredetected, and an upper envelope wasestimated by interpolation;thelowerenvelopewasestimatedinananalo- gousmannerfromthelocalminima(Fig.1,bottom).Finally, theaveragedifferencebetweenbothenvelopesalongtime

wascomputed as an estimateof the f-wave meanampli- tudeintheleadexamined.Themathematicaldetailsofthis amplitudemeasurementalgorithmhavebeendescribedby Meoetal.[15].

Ofthesixfrontalleads,onlytwoprovidelinearlyinde- pendent voltages [16]. Hence, to avoid redundancy, only leadsI,IIandV1—V6wereconsideredinsubsequentanaly- ses.Amplitudecomputationwasperformed usingMATLAB, version2011a(MathWorks,Natick,MA,USA).

For each patient, the f-wave amplitude used for prediction was computed in every lead, using the available duration of atrial activity signal after TQ segment con- catenation. To validate the temporal stability of this measurementmethod,thef-wavemeanamplitudewasalso computed oninitial 10-secondand30-secondsegments of atrial activity in therecordings, wheresuch lengthswere available, and Pearson correlation coefﬁcients were then determinedforeverylead.

Univariate analysis

Distribution normality was ﬁrst checked for the variables under examination by the Kolmogorov—Smirnovtest.

Levene’s correction was applied when homoscedasticity (homogeneityofvariance)couldnotbeassumed.Underdata normality,groupswerecomparedbyaparametricStudent’s t-test,whereasanon-parametricMann—WhitneyU-testwas

Figure1. Thef-waveamplitudemeasurementalgorithmusedinthisstudy.Top:electrocardiogram(ECG)fromtheﬁrstpatientofthe database(toeasevisualization,only10secondsareshown);theTQintervals(dashedboxes)aresegmentedafterdetectingtheR-wave peaks(redcrosses)aswellastheQ-waveonsetandT-waveoffsetlocations.Middle:theTQsegments(dashedboxes)aremeancentred andconcatenated.Bottom:localmaxima(redcircles)andminima(bluecrosses)aredetectedintheconcatenatedTQsegments,andthen interpolatedtoyieldanestimateoftheupperandlowerenvelopes(redandbluedashedline,respectively)oftheatrialactivitysignal;

ateachtimeinstant,thedifferencebetweenthetwoenvelopesprovidesaninstantaneousestimateoff-waveamplitudeintheECGlead considered.

(5)

used when the variables did not show a normal distribution.Proportionanalysiswasbasedonthe␹²test.Foreach univariatepredictorofCAoutcome,receiveroperatingcharacteristic(ROC)curveswerecomputedtoﬁndthecut-off point providing the optimal trade-off between sensitivity and speciﬁcity, and the area under the ROC curve (AUC) wasusedasapredictionperformance index.This statisti- calanalysisandthelogisticregression(LR)modeldescribed nextwereperformedwithSPSSsoftware,version13.0(IBM, Armonk,NY,USA).

Multivariate analysis

An LR model was constructed from a linear combination of f-wave amplitudes computed in the ECG leads, acting asmultiplepredictorvariables.Optimallinearcombination coefﬁcientsweredeterminedbymaximum likelihoodesti- mationfromtheavailabledataset[17].Afterestimatingthe modelcoefﬁcients,theLRscorewascomputedfromthef- wave amplitudeset of each patient.The numericalvalue of the LR score is directly related to the estimated odds of CAsuccess (nota voltage),andwasthen usedasa CA outcome predictor.Usingthe LRscore,ROC-basedindices werederivedtoquantifythemultivariatemodelpredictive performance,asin theunivariate analysisabove.A back- wardelimination technique basedon theWald index [17]

wasemployedtoselecttheECGleadswhosef-waveampli- tudescontributedtotheLRpredictionscoreinastatistically signiﬁcantmanner.

Results

Study population

Sixty-two consecutive patients (52 men; mean age 61.5±10.4years)wereincludedinthestudy.Patientchar- acteristics are summarized in Table 1. Patients had a mean AF history of 61.5±56.1months. AF was persistent in54patients(87.1%)andlong-standingpersistentineight patients(12.9%).DurationoftheactualAFepisode(ongoing atthetimeofCA)was7.3±11.1months.AFwasidiopathic in26patients(41.9%).

Ablation procedure and follow-up

The flowchart of the ablation procedure is presented in Fig. 2. PV isolation was achieved in all patients. AF terminationby CAwasaccomplishedin23cases(37%), after a loading dose of amiodarone in three cases (5%) and by electrical cardioversion in 36 cases (58%). Procedure duration was 263.5±64.0minutes, fluoroscopy time was 14.1±6.3minutes and radiofrequency delivery time was 47.5±25.7minutes.Twomajorcomplicationsoccurred:one arteriovenous femoral fistula requiring surgery; and one intra-alveolarhaemorrhageprolonginghospitalization.

Five patients hada redo procedure for AF recurrence.

With 1.08 procedures per patient and a follow-up of 13.9±8.3months after the last procedure, AF recurred in 15 patients. Recurrences occurred after a delay of 5.5±4.2months. Four additional patients had an atrial tachycardia recurrence,for which they were successfully

Figure2. Flowchart ofthe ablationprocedurestepsand endpoints used in this study. AF: atrial ﬁbrillation; CA: catheter ablation;LA:leftatrium;PV:pulmonaryvein;RA:rightatrium.

reablated (classiﬁed in the CA success group). In the CA successgroup(n=47),40patientswereoffanti-arrhythmic drugsandsevenpatientswerestillonamiodarone,despite absenceof AFrecurrence(as pertheircardiologist’spres- cription).

Validation of f-wave measurement

Forthe34patientswithconcatenatedatrialsignalsegments ofover30seconds,Pearsoncorrelationoff-waveamplitude measuresusing10-secondand30-secondsegments wasat least0.98(P<0.001)inallleads.

Whencomparingtheamplitudecomputedovertheorig- inalatrial signal(actual lengthdependent onthe patient) with that over 10-second segments for the same patient population, the minimum correlation remained at 0.97 (P<0.001),keepingatleastat0.9inallleadsforsegments asshortas5seconds.Inaddition,theminimumcorrelation betweenamplitudes measuredininitialandﬁnal5-second segmentswas0.799,obtainedinleadV1.Fulldetailsofthe validationstudycanbefoundin[18].

Univariate analysis

Table 1 shows the results of the univariate analysis per- formedontheusualclinicaldata.OnlyAFhistoryduration presented a signiﬁcant intergroup difference, with the highestAUC(54.7±53.5 vs.87.0±61.0months;P=0.047;

(6)

Table1 Patientcharacteristicsfortheoverallpopulation,andcomparisonbetweentheCAsuccessandCAfailuregroups byunivariateanalysis.

Patientcharacteristics Overall (n=62)

CAsuccess (n=47;75.8%)

CAfailure (n=15;24.2%)

P

Men 52(83.9) 40(85.1) 12(80.0) 0.64^a

Age(years) 61.5±10.4 60.3±10.4 65.1±9.7 0.12^b

Hypertension 22(35.5) 19(40.4) 3(20.0) 0.15^a

Sleepapnoeasyndrome 5(8.1) 4(8.5) 1(6.7) 0.82^a

Diabetes 8(12.9) 8(17.0) 0(0) 0.09^a

Bodymassindex 27.6±4.4 27.6±4.6 27.8±3.7 0.90^b

Obesity 14(22.6) 10(21.3) 4(26.7) 0.66^a

Coronaryheartdisease 7(11.3) 7(14.9) 0(0) 0.11^a

Hypertensivecardiomyopathy 6(9.7) 2(4.3) 4(26.7) 0.011^a,d

Dilatedcardiomyopathy 8(12.9) 4(8.5) 4(26.7) 0.07^a

Valvularheartdisease 4(6.5) 2(4.3) 2(13.3) 0.21^a

Hypertrophiccardiomyopathy 2(3.2) 2(4.3) 0(0) 0.42^a

AFhistory(months) 61.5±56.1 54.7±53.5 87.0±61.0 0.047^c,d

CurrentAFepisode 7.3±11.1 6.8±12.0 8.8±7.3 0.06^c

LAanteroposteriordiameter(mm) 47.2±7.1 47.1±7.3 47.5±6.4 0.86^b

LAsurface(cm²) 29.1±5.9 28.4±5.8 31.5±5.5 0.09^b

LVEF(%) 59.4±15.6 61.5±13.5 53.1±19.7 0.07^b

LACTscanmaximalvolume(mL) 141.0±44.9 135.7±42.7 158.6±49.1 0.11^b

PatientsonamiodaronebeforeCA 26(41.9) 19(40.4) 7(46.7) 0.67^a

Patientswithrepeatprocedure 5(8.1) 4(8.5) 1(6.7) 0.82^a

Dataareexpressedasnumber(%)ormean±standarddeviation.CA:catheterablation;CT:computedtomography;LA:leftatrium;

LVEF:leftventricularejectionfraction.

aPvaluecomputedaccordingtothe␹²test.

b PvaluecomputedaccordingtoStudent’st-test.

c PvaluecomputedaccordingtotheMann—WhitneyU-test.

d Statisticalsigniﬁcance.

Table2 Results fromthe univariateanalysis onatrial ﬁbrillatory amplitudes computed in the electrocardio- gramleadsseparately.

Lead CAsuccess CAfailure P^a AUC I 30.76±23.27 43.1±26.13 0.039 0.678 II 50.87±31.75 51.57±21 0.38 0.576 V1 72.95±70.24 59.43±32.24 0.55 0.552 V2 81.30±194.9 74.69±70 0.36 0.579 V3 87.69±217.8 63.5±52.42 0.84 0.518 V4 78.7±169.1 55.57±55.29 0.93 0.508 V5 56.57±71.28 40.24±14.56 0.90 0.511 V6 40.93±41.87 33.99±14.08 0.88 0.513 Dataareexpressedasmean±standarddeviation(␮V)forboth groups of interest. AUC: area under the receiver operating characteristiccurve;CA:catheterablation.

aPvaluescomputedaccordingtotheMann—WhitneyU-test.

AUC=0.687).Thereweremorehypertensiveheartdisease patientsintheCAfailuregroup(4.3%vs.26.7%;P=0.01).

Nootherdifferencesbetweenthegroupsweresigniﬁcant.

Table2summarizestheunivariateanalysisresultsforthe f-waveamplitudes computedontheECG leadsseparately.

Owing to the high dispersion of the data, no signiﬁcant

differencesarefoundbetweengroupsfor anyleadconsid- ered on its own, except for lead I, with a maximal AUC value of 0.678. The correspondingROCcurve is displayed inFig.3,resultingin66%sensitivity, 73.3%speciﬁcityand 67.7%accuracyattheoptimalamplitudecut-offof0.029mV.

Multivariate analysis

AfterapplyingLRwithbackwardeliminationontheavailable clinical data (Table 1), no combination of clinical varia- bleswasretained.Concerningthef-waveamplitudevalues, a total of ﬁve iterations of the same LR with backward elimination protocol were run beforeall remaining variables reached the signiﬁcance level (P<0.05). The leads selected were: I (Wald index=6.882; P=0.009), V1 (Wald index=5.311; P=0.021),V2(Waldindex=6.826;P=0.009) and V5 (Wald index=5.038; P=0.025). When clinical and ECGdataweretakentogether,themodelselectedthesame remainingvariables.

Fig.4 illustratestheROCof theLRscorebasedonthe fourselectedleadsresultinginthebestCApredictionper- formance.Theoptimalcut-offvaluewas1.732,yielding83%

sensitivity,73.3%speciﬁcityand80.6%accuracy,assumma- rizedin Table3.The LRscoreandthus itsoptimalcut-off value arenotexpressedin mV,butasthe estimatedodds

(7)

Figure3. Receiver operatingcharacteristiccurve oftheﬁbril- latoryamplitudecomputedinleadI,yieldingthebestprediction amongf-waveamplitudesmeasuredinelectrocardiogramleadssep- arately(univariateanalysis).Arrowindicatesoptimalcut-offpoint.

AUC: areaunderthereceiveroperating characteristiccurve;CI:

conﬁdenceinterval.

ofCAsuccessaccordingtothemultivariatemodelbasedon f-waveamplitudes(seemethodssection).

Discussion

This work shows that the simultaneousanalysis of f-wave meanamplitude,automaticallycomputedin multipleECG leads, improves the prediction of CA long-term outcome comparedwithsingle-leadamplitudemeasuresinpatients withpersistentAF.OuranalysispointstoI,V1,V2andV5 asthe optimal leads for CA outcome predictionbased on f-waveamplitude. The LRscore derived fromtheseleads predictsCAsuccesswithover80%accuracy.Withtheclinical goalofpredictingtherapyoutcomebeforeCAreferral,and thus avoiding unnecessary procedures, we focused exclu- sivelyonvariablesacquired beforetheinterventionasthe mostsuitableforpatientselection.

Figure4. Receiveroperatingcharacteristiccurveofthelogistic regressionscorebasedonf-waveamplitudescomputedinleadsI, V1,V2and V5,yieldingoptimalperformance inmutivariatepre- dictionanalysis.Arrowindicatesoptimalcut-offpoint.AUC:area underthereceiveroperatingcharacteristiccurve;CI:conﬁdence interval.

F-wave amplitude as a predictor of ablation outcome

Avarietyofclinicalandsignalfeatureshavebeenexplored as predictors of CA outcome in the literature [3—9,15].

Frequency-domain analysis has received particular atten- tion,especiallytheAFcyclelength,whichislinkedtothe atrialmyocytes’refractoryperiod[19],andhasbeenasso- ciatedwithacuteAFterminationbyCA[5,9].TheAFcycle length’slink withsinus rhythm maintenance after CAhas alsobeendemonstrated[9],buthaslackedreproducibility infurther studies [5,20,21],thus questioning the spectral featuresasmarkersoflong-termCAoutcome.Toillustrate thislimitation,theAFdominantfrequencyinleadV1yielded an AUC of 0.662 (P=0.06) in our database. Similarly, the phase-lock index [10] was not predictive of CA outcome inourstudy(AUC=0.536;P=0.68),probablybecausethat indexhasbeenproposedinthecontextofaspeciﬁcablation

Table3 Confusionmatrixoftheﬁnallogisticregressionmodelwithoptimalcut-off.

Positiveoutcome (CAsuccess)

Negativeoutcome (CAfailure) Predictedpositiveoutcome(CAsuccess) True

positives=39

Falsepositives=4 PPV=90.7%

Predictednegativeoutcome(CAfailure) False negatives=8

Truenegatives=11 NPV=57.9%

Sensitivity=83.0% Speciﬁcity=73.3% Accuracy=80.6%

NPV:negativepredictivevalue;PPV:positivepredictivevalue.

(8)

procedure,aimingatthesuppressionoffocalimpulseand rotormodulationsources[22].

F-waveamplitude onthe surfaceECG isrelatedtothe magnitudeof the underlying atrial voltage, which is also relatedtotheamount ofviable atrialmuscle[6,7];ithas beenassociatedwithAFdurationandpatientage,bothin V1andII,withhighervalueslinkedtoAFacutetermination byCA[7].However,anatomicalcharacteristics,suchasLA sizeandleftventricularend-diastolicdiameterdonotseem tocorrelateconsistentlywithf-waveamplitudeindifferent leads[7];alackofcorrelationbetweenLAanatomyandCA outcomeis observed in ourstudy (Table 1).Similarly, the long-termpredictionabilityoff-waveamplitudemeasured inasingleleadhassofarremainedtenuous,orhasonlybeen demonstratedwithlimitedaccuracy,dependingonthelead considered [6,7]. Electrical dipoles associated with atrial activationwavefrontsduringAFmaycancelout,asaresult oflocal interactions,andcancellationeffects mayreﬂect differentlydependingontheECGlead,whichmayexplain thespatialvariabilityoff-waveamplitude.Moreover,spa- tialvariabilityislikelytoincreaseinpersistentformsofAF, whereatrialelectricalactivationbecomesmorecomplex.As anillustrationoftheconsequencesof thishypothesis,the amplitudeofleadIinTable2ishigherforfailingablations, whichisinconsistentwiththeexpectedresults,pointingout thelackofreliabilityofsingle-leadmeasures,andconﬁrm- ingthevalueofconsideringmultipleleads,asinthepresent approach.

The algorithm for f-wave amplitude computation describedinthemethods sectiononly considerstheatrial signal duringthe TQintervals, asinthe studies byCheng etal.andNault etal.[6,7],but, unlikethoseworks,itis notmanual.Also,incontrasttothestudybyChengetal.[6], ourestimationprocedure doesnotrequireanintracardiac EGMinputastimereference,andisthusfullynon-invasive.

Thehighcorrelationresultsinourstudyshowthegoodtem- poralstabilityandreproducibilityoftheautomatedf-wave amplitudemeasures. This study focused onan automated computation of f-wave amplitude in order to avoid the subjectivity of human operators in manual measurement methods requiring visualinspection. Hence, a comparison withmanualmethodswasnotconsideredrelevant.

Multivariate analysis of ablation outcome

MostoftheexistingworksaimingatCAoutcomeprediction considereach variableseparatelyfromtherest by means of univariate analyses, and the whole set of variables is onlyintroduced intoa multivariate model,such asLR,to identify independent predictors [6,7,9]. By contrast, the presentworkdirectlyexploitstheLRscoreasaprediction index,basedonthemultiplevariablesincorporatedintothe model. Abackward elimination procedure is employed to selectthemost discriminantvariablesin thismultivariate setting.Byconsidering multiple ECGleads simultaneously, thisapproachisabletostressthelong-termoutcomepre- dictioncapabilitiesofthef-waveamplitudeinthecontextof theCAtherapy,asshowninTable3,Figs.3and4.Suchlong- termpredictioncapabilitiesaresuperiortothosepreviously reported[6,7].OptimallycombiningECGleadsthroughsuit- ablesignaldecompositiontechniqueshasalsorecentlybeen proposedbyMeoetal.[15].Althoughthatapproachyielded

improvedpredictionofacuteCAoutcome,itdidnotprovide satisfactoryresultsinthelong-termpredictiongoalofthe presentinvestigation.

Surface ECG lead selection

Typically,only onelead (e.g.,II, aVFor V1) isconsidered separately,withoutexploitingtheintrinsicspatialvariabil- ityofthef-wavesacrossECGleads[6,7,9].Themultivariate LRmodelusedinthisworkpointstoleadsI,V1,V2andV5 asthemostdiscriminanttodeterminemid- andlong-term CAoutcomebasedonf-waveamplitude,withaprediction accuracyexceeding80%,whilekeepingacceptablevaluesof specificityandsensitivity.Tosupporttheconsistencyofour analysis,analternativefeatureselectionapproachbasedon supportvectormachines[23]wasalsotested,retainingthe samesetofleads,andyieldingequivalentpredictionresults evenaftercross-validation.Theseresults stressthebene- fitsoftakingintoconsiderationtheinformationfromseveral leads thatareoften neglectedin AFanalysis.Byvirtueof theirspatiallocation,theselectedleadsV2andV5mayoffer amoredistinctiveviewofLAactivity,whichmaybeinsuf- ficientlyrepresentedbyV1,yetcouldplayasignificantrole inAFcharacterizationandpredictionoftherapyoutcome.

Clinical implications

ProposingCAtoapatientwithpersistentAFisadifficultclin- icaldecision.Currentguidelines[2]statethattheindication classforpersistentformsofAFisweakerthanforparoxysmal AF.SuboptimalresultswithCAarenotoriousforsuchforms, andtheablationtechniquestillexposespatientstoasignifi- cantriskofmajorcomplications[24].Thisdecisionmayrely onlyonclinicalandimaginginformation,butthemostvalu- ablepredictors ofCA results(suchasmagnetic resonance imaging-basedquantificationofatrialfibrosis[25])arenot availableinthemajorityofelectrophysiologylaboratories.

Hence,developingastandard12-leadECG-basedpredic- torofCAoutcomeinthecontextofpersistentAFmayhave majorclinical value, byproviding arapid toolfor patient selection. Softwaremaybeeasily integratedinECGs,and canprovideaninstantaneousmultileadcomputationofthe f-waveamplitudethatcliniciansmayuseforﬁne-tuningthe clinicalindicationforablation.

Study limitations

Although themultivariateclassiﬁcationandleadselection approaches(LR,supportvectormachines)consideredinthe analysisareincloseagreement,theirgeneralizationability maybehamperedbytherelativelysmallsamplesizeavail- ableinourstudy.Thissamplesize,however,issimilartothat insimilarstudiesrecentlyreportedintheliterature[6,20].

Thestepwiseablationprocedureimpliesacertaindegree ofinterpatientvariability,whichmayrepresentabiasforCA outcome.Nevertheless, amongthedescribed strategiesof persistentAFablation,thestepwiseapproachisadoptedin manyablationcentres,andencompasseswidespreadstrate- gies,suchascircumferentialPVisolationaloneandisolated CFAE ablation [5—7,9,20]. While electrical cardioversion might be considered asa confoundingfactor when evalu- atingCAoutcome,itsabilitytomaintainsinusrhythmwhen

(9)

appliedonitsowntotreatpersistentAFremainscontrover- sial[26].Inanycase,theproportionsofpatientswhohad receivedcardioversionintheCAsuccessandfailuregroups ofourdatabasewerenotsigniﬁcantlydifferent.Likewise, the continueduse of amiodaronebeforeablation and the applicationofrepeatproceduresdonotseemtoinﬂuence CAoutcome(Table1).Thus,weconsiderthattheprediction valueofthecomputationalmethodonthesurfaceECGmay beextrapolatedwidelyinclinicalpractice.

AFterminationrateduringCAwasinferiorin ourstudy compared with some published series [27,28]. Neverthe- less, asstatedin the methods section,AF terminationby ablationwasnotanendpoint,andoperatorsperformedlim- itedradiofrequencydeliveriesoutsidethePVantra;thisdid not hamperthe long-term outcome of CA in this formof AF,whichis inaccordancewiththerecentliterature[29].

Indeed,AFtermination byablation isdebatable, andsev- eralserieshaveshownnopredictivevalueofAFtermination duringablationforlong-termoutcome[30].Thebestchoice ofextravenousablationtargetsinpersistentAFiscurrently beinginvestigated,withdifferentconceptsandtechnologies otherthanCFAEablation,butwithoutconsensusforthetime being[31,32].

Anti-arrhythmic drugs were continued in a minorityof patients in the CA success group, without evidence of arrhythmia recurrence. As follow-up was performed in a

‘‘real-life’’ setting, drug prescription was continued as deemednecessary bythecardiologisttreatingtherespec- tivepatients. We considerthepotentially induced biasas limited, as only a few patients were concerned, and the same drugs had been ineffective in these same patients beforetheablationprocedure.Clinicalimprovementbeing acertaintyinthesepatients,theresultsreﬂectCAoutcome inan‘‘intention-to-treat’’real-lifesettingevenbetter.

Follow-upwasperformedwithsystematic24-hourHolter ECG monitoring (and additional ECG or Holter in case of symptoms),accordingtocurrent guidelines[2],asinsim- ilar studies[6,7,9]. However,the recurrenceratemay be underestimatedcomparedwithimplantablelooprecorders.

Finally,asinother predictionstudies(e.g.[5—7,9,10]), ourresultsarebasedonretrospectivedata,andshouldbe conﬁrmedbyfurtherprospectiveanalyses.

Conclusions

SurfaceECGrecordingsprovideusefulnon-invasiveinforma- tionforidentifyingpersistentAFpositiveresponderstoCA.

Thisworkhasshownthatconsideringthef-waveamplitude across multiple ECG leads jointly yields improved long- term outcome prediction compared with single-channel amplitudemeasures.Thenovelpredictorenhancespatient selection for CA,and can thus contribute tothe patient- tailoredtreatmentofAF.

Sources of funding

This work was supported in part by the French National Research Agency (contract ANR-2010-JCJC-0303-01

‘‘PERSIST’’) and the Scientiﬁc Centre of Monaco.

A.R.H.-M. was in receipt of a postdoctoral research

fellowship awarded by the University of Nice Sophia Antipolis.V.Z.isamemberoftheInstitutUniversitairede France.

Disclosure of interest

Theauthorsdeclarethattheyhavenocompetinginterest.

References

[1]January CT, Wann LS, Alpert JS,et al. 2014 AHA/ACC/HRS Guideline forthe managementofpatientswithatrial ﬁbril- lation:executivesummary:areportoftheAmericanCollege ofCardiology/AmericanHeartAssociationTaskForceonPrac- ticeGuidelinesandtheHeartRhythmSociety.JAmCollCardiol 2014;64:2246—80.

[2]CalkinsH,KuckKH,Cappato R,etal.2012HRS/EHRA/ECAS Expertconsensusstatementoncatheterandsurgicalablation ofatrialﬁbrillation: recommendationsforpatientselection, procedural techniques,patient management and follow-up, deﬁnitions, endpoints, and research trial design. Europace 2012;14:528—606.

[3]D’Ascenzo F, Corleto A, Biondi-Zoccai G, et al. Which are themostreliablepredictorsofrecurrenceofatrialﬁbrillation aftertranscatheter ablation?:ameta-analysis. IntJCardiol 2013;167:1984—9.

[4]GanesanA,ShippN,BrooksA,etal.Long-termoutcomesof catheter ablationof atrial ﬁbrillation: a systematic review.

HeartLungCirc2012;21:S130[abstract318].

[5]HeistEK,ChalhoubF,BarrettC,DanikS,RuskinJN,Mansour M.Predictorsofatrialﬁbrillationterminationandclinicalsuc- cessofcatheterablationofpersistentatrialﬁbrillation.AmJ Cardiol2012;110:545—51.

[6]ChengZ,Deng H,ChengK,et al.Theamplitude ofﬁbrilla- torywavesonleadsaVFandV1predictingtherecurrenceof persistentatrialﬁbrillationpatientswhounderwentcatheter ablation.AnnNoninvasiveElectrocardiol2013;18:352—8.

[7]NaultI,LelloucheN,MatsuoS,etal.Clinicalvalueofﬁbrilla- torywaveamplitudeonsurfaceECGinpatientswithpersistent atrialﬁbrillation.JIntervCardElectrophysiol2009;26:11—9.

[8]ButtuA,VanZaenJ,VisoA,etal.Terminationofatrialﬁbril- lationbycatheterablationcanbesuccessfullypredictedfrom baselineECG.ComputCardiol2012;39:189—92.

[9]MatsuoS,LelloucheN,WrightM,etal.Clinicalpredictorsof terminationandclinicaloutcomeofcatheterablationforper- sistentatrialﬁbrillation.JAmCollCardiol2009;54:788—95.

[10]Jones AR, Krummen DE, Narayan SM. Non-invasive identifi- cation of stable rotors and focal sources for human atrial fibrillation:mechanisticclassificationofatrialfibrillationfrom theelectrocardiogram.Europace2013;15:1249—58.

[11]Squara F,LatcuDG, MassaadY, Mahjoub M, BunSS, Saoudi N. Contact forceand force-time integralin atrial radiofrequency ablation predict transmurality of lesions. Europace 2014;16:660—7.

[12]NademaneeK,McKenzieJ,KosarE,etal.Anewapproachfor catheterablationofatrialﬁbrillation:mappingoftheelectro- physiologicsubstrate.JAmCollCardiol2004;43:2044—53.

[13]PanJ,TompkinsWJ.Areal-timeQRSdetectionalgorithm.IEEE TransBiomedEng1985;32:230—6.

[14]CabassonA,MesteO.Timedelayestimation:anewinsightinto theWoody’smethod.IEEESignProcLett2008;15:573—6.

[15]MeoM,ZarzosoV,MesteO,LatcuDG,SaoudiN.Spatialvari- abilityof the12-leadsurface ECGasa toolfor noninvasive prediction ofcatheterablation outcome inpersistent atrial ﬁbrillation.IEEETransBiomedEng2013;60:20—7.

(10)

[16]Malmivuo J,Plonsey R. Bioelectromagnetism: Principlesand ApplicationsofBioelectricandBiomagneticFields.NewYork:

OxfordUniversityPress;1995.

[17]Indrayan A. MedicalBiostatistics.Chapman & Hall/CRCBio- statisticsSeries.BocaRaton,FL:CRCPress;2012.

[18]MeoM,Hidalgo-Mu˜nozAR,ZarzosoV,MesteO,LatcuDG,Saoudi N.F-waveamplitudestabilityonmultipleelectrocardiogram leadsinatrialﬁbrillation.ComputCardiol2015;42:505—8.

[19]Bollmann A, Kanuru NK, McTeague KK, Walter PF, DeLurgio DB,LangbergJJ.Frequency analysisofhumanatrialﬁbrilla- tionusingthesurfaceelectrocardiogramanditsresponseto ibutilide.AmJCardiol1998;81:1439—45.

[20]Di Marco LY, Raine D, Bourke JP, Langley P. Characteris- tics of atrial ﬁbrillation cycle length predictrestoration of sinus rhythm by catheter ablation. Heart Rhythm 2013;10:

1303—10.

[21]GaribaldiM,ZarzosoV,LatcuDG,SaoudiN.Predictingcatheter ablationoutcome in persistent atrialﬁbrillation usingatrial dominantfrequencyandrelatedspectralfeatures.ConfProc IEEEEngMedBiolSoc2012;2012:613—6.

[22]Narayan SM, Krummen DE, Rappel WJ. Clinical mapping approach to diagnose electrical rotors and focal impulse sourcesforhumanatrialﬁbrillation.JCardiovascElectrophys- iol2012;23:447—54.

[23]GuyonI,WestonJ,BarnhillS,VapnikV.Geneselectionforcan- cerclassiﬁcationusingsupportvectormachines.MachineLearn 2002;46:389—422.

[24]DeshmukhA,PatelNJ,PantS,etal.In-hospitalcomplications associatedwithcatheterablationofatrialﬁbrillationinthe

UnitedStatesbetween2000and2010:analysisof93801pro- cedures.Circulation2013;128:2104—12.

[25]McLellanAJ,LingLH,AzzopardiS,et al.Diffuseventricular ﬁbrosismeasuredbyT(1)mappingoncardiacMRIpredictssuc- cessofcatheterablationforatrialﬁbrillation.CircArrhythm Electrophysiol2014;7:834—40.

[26]Abu-El-Haija B, Giudici MC. Predictors of long-term maintenance of normal sinus rhythm after successful electrical cardioversion.ClinCardiol2014;37:381—5.

[27]FaustinoM,PizziC,CapuzziD,etal.Impactofatrialﬁbrilla- tionterminationmodeduringcatheterablationprocedureon maintenanceofsinusrhythm.HeartRhythm2014;11:1528—35.

[28]Pascale P,Shah AJ, RotenL, et al. Pulmonary veins to left atrium cycle length gradient predicts procedural and clinical outcomes of persistent atrial ﬁbrillation ablation. Circ ArrhythmElectrophysiol2014;7:473—82.

[29]MontL,BisbalF,Hernandez-MadridA,etal.Catheterablation vs.antiarrhythmicdrugtreatmentofpersistentatrialﬁbrilla- tion:amulticentre,randomized,controlledtrial(SARAstudy).

EurHeartJ2014;35:501—7.

[30]YokoyamaY,HiraoK.Atrialﬁbrillation(AF)terminationbyabla- tioninpatientswithpersistentAF—whatisthatsupposedto mean?CircJ2013;77:1399—400.

[31]HaissaguerreM, Hocini M,DenisA, etal. Driverdomains in persistentatrialﬁbrillation.Circulation2014;130:508—30.

[32]SwarupV,BaykanerT,RostamianA,etal.Stabilityofrotorsand focalsourcesforhumanatrialﬁbrillation: focalimpulseand rotormapping(FIRM)ofAFsourcesandﬁbrillatoryconduction.

JCardiovascElectrophysiol2014;25:1284—92.