ebook img

Automatic Detection of Atrial Fibrillation and Other Arrhythmias in ECG Recordings Acquired by a PDF

14 Pages·2017·1.66 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Automatic Detection of Atrial Fibrillation and Other Arrhythmias in ECG Recordings Acquired by a

electronics Article Automatic Detection of Atrial Fibrillation and Other Arrhythmias in ECG Recordings Acquired by a Smartphone Device LuciaBilleci1,* ,MagdaCosti2,DavidLombardi2,FrancoChiarugi3andMaurizioVaranini1 1 InstituteofClinicalPhysiology,NationalResearchCouncilofItaly,viaMoruzzi1,56124Pisa,Italy; [email protected] 2 CardiolineS.p.A.,viaLinz19-21-21,38121SpinidiGardolo(TN),Italy;[email protected](M.C.); [email protected](D.L.) 3 DedalusS.p.A.ViaGaetanoMalasoma,24,56121Pisa,Italy;[email protected] * Correspondence:[email protected];Tel.:+39-050-315-2354 (cid:1)(cid:2)(cid:3)(cid:1)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:1) (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7) Received:9August2018;Accepted:14September2018;Published:16September2018 Abstract: Atrialfibrillation(AF)isthemostcommoncardiacdiseaseandisassociatedwithother cardiaccomplications. FewattemptshavebeenmadefordiscriminatingAFfromotherarrhythmias and noise. The aim of this study is to present a novel approach for such a classification in short ECG recordings acquired using a smartphone device. The implemented algorithm was tested on the Physionet Computing in Cardiology Challenge 2017 Database and, for the purpose of comparison,ontheMIT-BHAFdatabase. Afterfeatureextraction,thestepwiselineardiscriminant analysis for feature selection was used. The Least Square Support Vector Machine classifier was trainedandcross-validatedontheavailabledatasetoftheChallenge2017. Thebestperformance was obtained with a total of 30 features. The algorithm produced the following performance: F Normalrhythm=0.92;F AFrhythm: 0.82;F Otherrhythm=0.75;GlobalF =0.83,obtainingthe 1 1 1 1 thirdbestresultinthefollow-upphaseofthePhysionetChallenge. OntheMIT-BHADFdatabase the algorithm gave the following performance: F Normal rhythm = 0.98; F AF rhythm: 0.99; 1 1 Global F = 0.98. Since the algorithm reliably detect AF and other rhythms in smartphone ECG 1 recordings,itcouldbeappliedforpersonalhealthmonitoringsystems. Keywords: electrocardiogram;smartphone;atrialfibrillation;arrhythmias;supportvectormachine 1. Introduction Atrialfibrillation(AF)isthemostcommoncardiacarrhythmiaandisassociatedwithseveral complications. AccordingtotheGlobalBurdenofDisease, theprevalenceofAFisof33.5million persons,affecting2.5%to3.2%ofpopulationworldwide[1].Severalstudiessuggestarisingprevalence andincidenceofAF;itisexpectedthatAFwillaffect6–12millionpeopleintheUSAby2050and 17.9millioninEuropeby2060[2]. Possiblecausesofthisriseareenhanceddetection,increasingageing ofthepopulation,andlifestylechanges[3,4]. AFisassociatedwithincreasedcardiovascularproblems, especially stroke and heart failure, as well as increased mortality [5]. Thus, the detection of AF is helpfultopreventsubsequentcomplications. Academiaandindustryhaverespondedtothisneedwithanincreaseinresearchanddevelopment regardingdevicesandstudiestowidenthechoicesforscreeningandmonitoringAF. ThetraditionalcardiacmonitoringtechnologiesforAFdetection,liketheHolterdevice,haveone or more drawbacks that limit the application of signal monitoring outside the clinics, including laborious workflow, the need for specialized staff trained in arrhythmia detection, poor patient compliance,cost,andinvasiveness. Recently,ECGdatahavebeencollectedusingnoveldeveloped Electronics2018,7,199;doi:10.3390/electronics7090199 www.mdpi.com/journal/electronics Electronics2018,7,199 2of14 technologies,includingsingle-leadECGadhesivesensors,implantablelooprecorders,smartphone attachments,andwearables[6]. Recentadvancesinthesewearableandpointofcaredevicescould potentiallyallowthedetectionofAFepisodesoutsideofclinicalsettingsinaneasyandcost-effective way. Importantly, the computational power and the storage capacity of these devices allow the executionofcomplexalgorithmsinrealtime, makingpossibletheonlinedetectionofAForother cardiacepisodesusingstand-aloneapplicationsathome. Different approaches have been adopted so far for AF detection, mainly using traditional technologies. These approaches consist in inter-beat (RR) intervals analysis [7,8], time-frequency analysisofECG[9,10],analysisbasedonPwaveabsence(PWA)andfrequencyspectrumanalysis (FSA)[11,12]. AlgorithmsbasedonPwaveshavealowperformanceinthepresenceofnoiseasthese wavesarepronetocontaminationwithmotionandnoiseartefacts[13]. Thus,mostofthemorerecent approachesforAFdetectionarebasedonRRanalysis.However,somealgorithmshavebeendeveloped combiningRRanalysisandanalysisofPWAand/orFSA,toenhancedetectionperformance[14–16]. Moreover,toimproveaccuracyandspecificityinAFdetection,severalpatternrecognitionapproaches have been recently proposed including neural networks [17,18]. In particular, the Support Vector Machine(SVM),anon-parametricclassifier,hasbeenemployedasitgivespromisingresultsinvarious medicaldiagnostics[19]includingAFdetection[20,21]. Oneofthemajorlimitationsofthealgorithmsavailabletodateisthattheyareusuallyaimedonly atdistinguishingAFfromnormalrhythmsandnotfromtheotherarrhythmias. Veryfewattempts havebeenmadeinthisdirection. Carraraetal.[22]appliedaK-NN(K-NearestNeighborhood)to classifyAF,normalsinusrhythm(NSR)andsinusrhythm(SR)withfrequentectopy,usingentropy, localdynamics,andfractalscaling. Theauthorsachievedhighperformanceontheirowndatabaseof 2722patientswith24hECGrecordings(AF:97%,NSR:98%andSRwithectopy: 90%). Theauthors also tested the algorithm on MIT–BIH databases obtaining lower performance for SR with ectopy (40%)thussuggestingthatdynamicalmeasuresdidnotdistinguishatrialfromventricularectopy[23]. Entropy estimation was also applied to distinguish AF from ventricular tachycardia/ventricular fibrillationobtaininganaccuracyof98%[24]. Uptonow,noneofthepublishedstudieshavetesteda widesetofarrhythmiastoevaluatehowaccuratelytheycanbediscriminatedfromAF. Recently,thePhysionet/ComputinginCardiologyChallenge2017encouragedthedevelopment ofalgorithmstoclassifysingleshortECGleadrecordings,asNSR,AF,otherrhythm,ortoonoisytobe classified[25]. Notably,theserecordingsareobtainedusingapointofcaredevice,AliveCorTM,anovel smartphone connected ECG device coupled with an application to record and diagnose the ECG, whichwaspreviouslydemonstratedtoprovideanaccurateandsensitivesingle-leadECGdiagnosisof AF[26,27]. Inthispaper,weproposeanovelapproachfortheclassificationofsingle-leadshortECGrecording as AF, normal rhythms, other type of rhythms and noise, by combining the three most important characteristicsofAF,i.e.,RR,PWAandFSAfeatures. Amulti-classSVMwasthenapplied,forarobust classification.TheproposedapproachwastestedonthePhysionetComputinginCardiologyChallenge 2017database[28]. ThedevelopedcodewassubmittedtothePhysionetOpen-SourceChallengecall andisavailablefordownload(https://www.physionet.org/challenge/2017/sources/). Forthesake of comparison, the algorithm was also tested on the MIT-BIH AF database (AFDB), consisting in long-termECGsignalsacquiredusingambulatoryECGrecorders,althoughinthisdatabaseonlyAF andNSRareannotated. 2. MaterialsandMethods 2.1. Datasets The Physionet Computing in Cardiology Challenge 2017 database consists of a total of 12,186 ECG recordings. These data are single-channel signals obtained by the AliveCorTM Kardia Electronics2018,7,199 3of14 smartphone connected device. Data were bandpass filtered (0.5–40 Hz) and stored at 300 sps, with16bitprecisionanda±5mVdynamicrange. Electronics 2018, 7, x FOR PEER REVIEW 3 of 14 The dataset was divided in a dataset, available to the public, for the development, training, and validation of the algorithms and a hidden dataset, not available to the public, on which the The dataset was divided in a dataset, available to the public, for the development, training, and perforvmalaidnacteioon fotfh ethael gaolgroitrhitmhmssw aansde va ahluidadteend dbaytatsheet, Cnhoat llaevnagileaboler gtaon tihzee rsp.ubTlhice, oanv awilahbiclhe dthaet aset consispteedrfoorfm8a5n2c8er eocf otrhdei nalggsolraitshtminsg wfraosm ev9aslutaote6d1 sbyw thhiele Cthhealhleindgdee nordgaantaizseertsc. oTnhtea ianveadila3b6l5e8 draetcaosredt ings ofsimciolanrsilsetendg tohfs 8a5n2d8 crleacsosrddiinsgtrsi blausttiiongn sf.roTmhe 9r esc otord 6i1n gss wwheirlee ltahbe ehleidddaesn bdealotansgeti ncgontotaifnoeudr d36if5f8e rent classerse:cNorSdRin,gAs Fo,fo stihmeirlarrh lyetnhgmth,s aanndd ncloaisssy driesctroibrdutiinognss.. AThftee rretchoerdoinffigcsi awleCreh laalbleenlegde acsl obseulorneg,itnhge rteo was four different classes: NSR, AF, other rhythm, and noisy recordings. After the official Challenge follow-up phase in which the annotations were reviewed and re-evaluated by experts to create a closure, there was follow-up phase in which the annotations were reviewed and re-evaluated by correctedversion(v3). FurtherdetailsaboutthePhysionetComputinginCardiologyChallenge2017 experts to create a corrected version (v3). Further details about the Physionet Computing in databasecanbefoundinCliffordetal.[25]. Cardiology Challenge 2017 database can be found in Clifford et al. [25]. TheAFDBdatabase[29]includes25fullyannotatedECGrecordingsandatotalof299AFepisodes The AFDB database [29] includes 25 fully annotated ECG recordings and a total of 299 AF (mostlyparoxysmal). EachECGrecordingisapproximately10hlongandissampledat250Hz. Italso episodes (mostly paroxysmal). Each ECG recording is approximately 10 h long and is sampled at 250 contaiHnzs. sIot malseoc caosnetsaionfs astormiael cflausettse orfa antrdiaNl fSluRt.ter and NSR. InFigIun rFeig1utrhe e1 ftuhlel fpurlol pcerossceosfs tohf ethper porpoopsoesdedm meeththoodd iiss rreeppoorrtteedd. . ECG Signals acquired using Alivecordevice SVM Classification Learning: ten-fold cross validation on 80% of the available dataset Feature extraction Feature selection Pre-processing RR, PWA, FSA Stepwise discriminant analysis Validation: 20% of the available dataset RR series extraction and correction Normal rhythms AF rhythms Other rhythms Noise Figure1.Flowchartoftheproposedalgorithm.Abbreviationsare:Pwaveabsence(PWA),frequency Figure 1. Flowchart of the proposed algorithm. Abbreviations are: P wave absence (PWA), frequency spectrsupmectarnuamly asniasl(yFsiSsA (F),SaAt)r,i aatlrfiiabl rfiibllrailtliaotnion(A (AF)F.). 2.2. ECGPre-ProcessingandRRSeriesConstruction 2.2. ECG pre-Processing and RR Series Construction ECGEsCigGn asligsnwales rweefirres ftirssut bsmubimttietdtedto tos osmomee pprree--pprroocceessssiinngg stsetpeps sini norodredr etro troedruecdeu ncoeisneo ainsed and improivmepqrouvael iqtyu.aFliitrys.t F,iirmstp, iumlspiuvlesiavret iafratciftaccat ncacneclienligngw wasaso obbtataiinneedd bbyy ccoommppaarriningg ththe eECECG Gsigsniganl awliwth itha mediaan mfieldteiarne dfilsteigrenda sli(g6n0alm (6s0w misn wdoinwdo):wt)h: ethEeC EGCGv avlaulueess wwhhoossee aabbssoolulutet edidffiefrfeenrecne cweitwh itthhe tfhilteerfieldte red oneseoxnceese edxecededaetdh rae tshhroesldhowlde wreerree prelpaclaecdedw witihtht thhee aavveerraaggee oof fththe evavlauleuse bsebfoerfeo arenda nafdtear fttheermth [e3m0]. [30]. The resulting signal was then upsampled to 1200 Hz to increase the accuracy in the localization of Theresultingsignalwasthenupsampledto1200Hztoincreasetheaccuracyinthelocalizationof QRSs thus reducing the quantization of the RR intervals. QRSsthusreducingthequantizationoftheRRintervals. QRS detection was performed by applying a threshold on the absolute amplitude of a filtered QRSdetectionwasperformedbyapplyingathresholdontheabsoluteamplitudeofafiltered derivative signal. This threshold was updated at each new detection and was changed with the derivative signal. This threshold was updated at each new detection and was changed with the temporal distance from the previous QRS. The fiducial point of each QRS was selected as the time temporaldistancefromthepreviousQRS.ThefiducialpointofeachQRSwasselectedasthetime occurrence of the maximum (minimum) of the signed derivative signal. The beginning and the end occurroefn QceRSo fwtehree mesatixmimatuedm b(ym thinei mcrousmsin)go fotfh deersiivganteivde dtherroivuagthiv tehes i0g.2n5a lt.hTrehsehobledg. iTnhniisn aglgaonridthtmhe foern dof QRSwQeRrSe deestteimctiaotne dhabsy bteheen cprroesvsioinugsloy fpduebrliisvhaetdiv beyt horuor uggrohutph ean0d.2 o5btthairneesdh oal dve.rTyh giosoadlg poerriftohrmmafnocreQ RS detect[i3o1n]. hTahse RbeRe snerpierse wviaosu tshleyn poubtbaliinsehde das bthyeo duifrfegrreonucep baentwdeoenb tsauicnceedssiavev eQrRySgs.o odperformance[31]. TheRRserieswasthenobtainedasthedifferencebetweensuccessiveQRSs. 2.3. Feature Extraction and Selection 2.3. FeatureExtractionandSelection The extracted features can be categorized in four types: (1) derived from the RR series analysis; (2) derived from PWA; (3) derived from FSA; and (4) other features derived from the ECG and QRS Theextractedfeaturescanbecategorizedinfourtypes: (1)derivedfromtheRRseriesanalysis; analysis. A set of 55 features was considered. The motivation for the selection of these features is that (2)derivedfromPWA;(3)derivedfromFSA;and(4)otherfeaturesderivedfromtheECGandQRS in clinical practice arrhythmia can be distinguished from normal heartbeats in terms of both Electronics2018,7,199 4of14 analysis. Asetof55featureswasconsidered. Themotivationfortheselectionofthesefeaturesis that in clinical practice arrhythmia can be distinguished from normal heartbeats in terms of both morphologicalanddynamicdifferences. Indeed,arrhythmiaaretypicallyassociatedwithvarious irregularitiesinheartrhythm,butcanusuallybecharacterizedalsobydifferentabnormalordistorted patternsinthewaveformshape(e.g.,distortedQRScomplex)orbythemmissingsomeimportant components (e.g., P wave). For this reason, we decided to use a combination of different types of featuresfortherepresentationoftherhythm. ANOVA was used for feature pre-selection and tuning (some features were log-transformed in order to normalize their distribution and increase their discriminant power), then the stepwise discriminantanalysiswithRao’Vcriterionforfeatureinclusionwasappliedtoselecttensetswithan increasingnumberoffeatures. ThefinalsetoffeatureswasdeterminedbythevalidationofLS-SVM classifiersasdescribedinSection2.4. The30finallyselectedfeatures,aswellastheirdiscriminant power(F),arereportedinTable1. Inthefollowing,agenericdescriptionofthefeaturesisreported, fordetailsseethesubmittedcode. Table1.Descriptionoftheselectedfeatures.F:discriminantpower. Feature F Description FeaturesextractedfromtheRRseries bigem 910 Bigemyrhythmindex trigem 440 Trigemyrhythmindex RRmean 250 MeanvalueoftheRRintervals RRmax 168 MaxvalueoftheRRintervals RRbrady 125 MeanvalueoftheRRintervalsabovethethresholdof1.2s RRtachy 659 MeanvalueoftheRRintervalsbelowthethresholdof0.7s CoSEn 1913 Coefficientofsampleentropy Mawsdo 1771 Meanoftheabsoluteweightedsuccessivedifferencewithweightoptimization Meanoftheabsoluteweightedsuccessivedifferencecontrollingforprematurebeats Mawsdco 2114 withcompensatorypause(weightoptimization) KFD 993 KatzFractalDimension FrequencySpectrumAnalysis(FSA) Poweraroundthepeakoftheweightedspectraldensityinthe3–20Hzband. mPdAF 357 Thepowerofthethreeharmonicswasconsidered mPdAFr 449 Ratiobetweenthepowerofthepeakandthepowerinthearound Pwaveanalysis(PWA) IndexofPwaveexistencebasedonthecomparisonbetweenamovingaverageof P_numr 88 0.06swidthandamovingaverageof0.12swidth IndexofPwaveexistenceobtainedcomparingamovingaverageof0.06swidthwith Pexist 67 theoutputofacombfilterconsistingoftwomovingaverageof0.04swidth, delayedby0.04s IndexofPwaveexistenceobtainedcalculatingthepowerpeakswithinintervals PintNr 240 beforetheQRSonsetandcomputingthemodeofthedistributionofthepositionsof suchpeaks OtherfeaturesextractedfromtheECGandtheQRScomplex Powerofthesignalobtainedasthedifferencebetweentheoriginalandthelow-pass pRhf 62 filtered(30Hz)signal(indexofnoise) QRSwidthMeanQ 1101 Trimmedmean(2%taildiscard)oftheQRSwidthvalues QRSwidthStdQ 545 Trimmedstandarddeviation(2%taildiscard)oftheQRSwidthvalues QRSwidthMax3Q 619 MeanofthethreemaxvaluesoftheQRSwidth QRSwidthMaxQs 292 Maxvalueofthemovingaverageon3valuesoftheQRSwidth QRSppAmplMea 44 Trimmedmean(2%taildiscard)oftheQRSpeak-peakamplitude RatiobetweentheQRSppAmplMeaandthepeak-peakamplitudeofthesignaloutside QRSpprAmplMea 311 theQRSTinterval Trimmedmean(3%taildiscard)oftheQRSpolaritycomputedasthesumofthe QRSpolMea 26 maximumandtheminimumoftheQRSvalues MinvalueoftheQRSequivalentwidths,definedastheQRSabsoluteareadividedby QRSwidthMin 183 thepeak-peakamplitude QRSwidthNmea 508 Trimmedmean(discarding40%ofhighvalues)oftheQRSequivalentwidths PnWideQRS 149 RatiobetweenthenumberofQRSwithequivalentwidth>1.25×QRSwidthNmea morphDiffr 550 MeanofthesumofabsolutedifferencesbetweentheQRSsandtheQRSmean QRSsadV 19 Sumoftheabsolutedifferencesbetweenthetwofirstright-singular-vectors nPVCp 353 Fractionoftheprematurebeatsthatarefollowedbyacompensatorypause AtypBeatPrk 450 RatiobetweentherelativeQRSwidthandthebeatprematurity Electronics2018,7,199 5of14 2.3.1. RRAnalysis ThefeaturesextractedfromtheRRserieswerethemostdiscriminant. Thesubsetselectedby step-wisediscriminantanalysisincluded: themeanandthemaxvalueoftheRRintervals;anindexof tachycardia;anindexofbradycardia;andindexofbigeminyandanindexoftrigeminy;themeanof theabsoluteweightedsuccessivedifference(Mawsd);thecoefficientofsampleentropy(CoSEn)[32] andtheKatzFractalDimension(KFD)[33]. Forthecomputationofthebigeminyrhythmindex,wedefinedRRoddasthemeanofRRwith oddindexandRRevenasthemeanofRRwithevenindex. Thebigeminyindexwasthencalculatedas theabsolutevalueoftheratio: (RRodd−RReven)/(RRodd+RReven). Forthecomputationofthetrigeminyrhythmindex,wedefinedRRmdasthemedianfiltered RRusingawindowlengthof3values. Trigeminyindexwascomputedastheratiobetweenthesum oftheabsolutevalueofsuccessivedifferencesofRRandthesumoftheabsolutevalueofsuccessive differencesofRRmd. ThefeatureMawsdwasbuiltwiththeaimofgettinganindexwhichtakesintoaccountmostly theRRvariabilityduetoAF,excludingthatduetorespiratorysinusarrhythmiaandthatduetoother arrhythmicevents. Maswdwascalculatedasthemeanofabsoluteweightedsuccessivedifferences ofRRobtainedasfollows: (a)onlytheRRsuccessivedifferenceslowerthanahighthresholdvalue andhigherthanalowthresholdvaluewereconsidered(thesethresholdsareproportionalthemedian RR);(b)adifferentweightwasthenassignedtotheseRRdifferencesdependingonthesignchange betweentwoconsecutiveRRdifferences. Moreover,aspecificcontrolaimedatexcludingthechanges duetoprematurecontractionswithacompensatorypausewasintroducedinamodifiedversionof thefeature(Mawsdc). Wealsocomputedanoptimizedversionofthesetwofeatures(Mawsdoand Mawsdco). In this case, the initial values of the weights and thresholds which control the routine behavior, were optimized by the Mesh Adaptive Direct Search (MADS) algorithm. We used the toolbox“NOMAD”[34]withthetargettomaximizetheF scoreinAFdetectionusingasinglefeature 1 decisionrule. 2.3.2. PWaveAnalysis The existence of a P wave was evaluated by applying a previously published algorithm [35]. Thealgorithmisbasedontwomovingaveragefilters(widths:0.06sand0.12s),followedbyadynamic event-relatedthreshold. Amodifiedversionofthisalgorithmwasalsoapplied,whichusedamoving averageof0.06swidthandacombfilterconsistingoftwomovingaveragesof0.04swidthdelayed by0.04s. Inaddition,anindexofPwaveexistencewasalsoobtainedbyevaluatingthepowerofthe signalinaspecificintervalbeforetheQRScomplex. 2.3.3. FrequencySpectrumAnalysis The measures for the detection of the weak oscillation which characterize AF were extracted byfrequencyspectrumanalysis. SincethehighQRSTpowermaycoverthisweakAFcomponent, aQRSTcancellingprocedurewasapplied. ThiswasbasedonapproximatingeachQRSTbyPrincipal ComponentmethodusingthepowerfulSingularValueDecomposition(SVD).Thesignalaroundeach QRSwasweightedbyatrapezoidalwindowandstoredinthecolumnsofthematrixX,whichwas decomposed by SVD. A matrix Xr was then rebuilt from the SVD decomposition using only the eigenvectorsassociatedtothegreatest2or3eigenvalues. Theseeigenvectorscontainonlythesignal componentswhicharepowerfulandsynchronous,thusthecolumnsofthematrixXrapproximate theoriginalQRST.Asignal,containingmainlythecomponentwithventricularorigin,wasobtained byunweightingtheestimatedQRSTsegmentsandconnectingthemwithastraightline. Thissignal wassubtractedfromtheoriginalECG,obtainingaresidualsignalwhichretainstheAFcomponent. Figure2shows,ontheleft,anexampleofQRSTcancellation(fromthetop:theoriginalsignal,theQRST estimated component, and the residual signal where the oscillatory AF is enhanced), on the right, Electronics2018,7,199 6of14 theestimateofthepowerspectraldensityobtainedbytheWelchmethodshowingthepeakduetothe AFcoEmlecptroonnices n20t1.8, 7, x FOR PEER REVIEW 6 of 14 ThemeasuresrelatedtoAFextractedwere: thepowerofthesignalintheband4–10Hz;theratio The measures related to AF extracted were: the power of the signal in the band 4–10 Hz; the ratio betweenthepowerin4–10Hzbandandthepoweroutsidethisband; thepoweraroundthepeak between the power in 4–10 Hz band and the power outside this band; the power around the peak of ofaweightedspectraldensityinthe3–20Hzband(thepowerofthreeharmonicswasconsidered); a weighted spectral density in the 3–20 Hz band (the power of three harmonics was considered); the therartaitoiob beettwweeeenn tthhee ppoowweer roof fthteh epepaeka aknadn tdhet hpoewpeorw areoruanrdo.u nd. 1 −30 mV) 0.5 G ( 0 −35 EC −0.5 QRST (mV) −00..5501 D (dB/Hz) −−4450 e PS −50 V) 0.2 gn (m 0.10 −55 resS −0.1 −60 15 16 17 18 19 20 0 5 10 15 20 25 30 time (s) Frequency (Hz) (a) (b) FigurFeig2u.rRe e2c. oRredcoArd0 0A9000:9Q0:R QSRTSTca cnacneclellalatitoionn bbyy SSiinngguullaarr VVaalulue eDDeceocmompopsiotisointi o(SnV(DS)V (Da)) a(nad) aPnodwePro wer spectsrpaelcdtreanl sditeyns(iPtyS D(P)SDof) roef sriedsuidaulasli gsinganlal( b(b).). eeQQRRSSTT:: eessttimimaatetde dQQRSRTS; Tr;esrSegsnS:g rne:sirdeusaidl usiaglnsail ganftaelr after SVDScaVnDc ecallnacteiollnat.ion. 2.3.4.2O.3t.4h.e OrtFheera Ftuearteusres FromFrtohme wthhe owlehoEleC EGCaGn ainn dinedxexo fofn nooisisee wwaass oobbttaaiinneedd aas sthteh epopwoewr eorf odfifdfeirfefnercee nbceetwbeeetnw teheen the original signal and the 30 Hz low-pass pass filtered ECG (pRhf). From the QRS complex, measures originalsignalandthe30Hzlow-passpassfilteredECG(pRhf). FromtheQRScomplex,measures of QRS width, peak-peak amplitude, and polarization were extracted. Other features were obtained ofQRSwidth,peak-peakamplitude,andpolarizationwereextracted. Otherfeatureswereobtained combining QRS morphology measures (e.g., QRS width) and rhythm (e.g., prematurity). Descriptions combiningQRSmorphologymeasures(e.g.,QRSwidth)andrhythm(e.g.,prematurity). Descriptions of these features are reported in Table 1. ofthesefeaturesarereportedinTable1. 2.4. Classification 2.4. Classification In this study, we applied the LS-SVM (Least Square SVM) classifier [36], which is derived by the Ionritghinisals StuVdMy ,[3w7]e. Aaps pLlSie-SdVtMh einLvSo-lSveVsM a l(eLaesta ssqtuSaqruesa rceosSt VfuMnc)ticolna swsiifithe req[u36al]i,tyw choincshtriasindtse,r tihvee d by theosroigluintiaolnS cVanM be[ 3o7b]t.aiAnesdL bSy- SsVolMvinign av osylvsteesma olfe laisnteasrq euqauraetsiocnoss itnf uthnec ttriaonnsfworimthede qspuaacleit. yDceofinnisntgra ints, thesoal uptoisointivcea nkebrneeol:b tainedbysolvingasystemoflinearequationsinthetransformedspace. Defining apositivekernel: k(cid:0)kx(cid:3435)𝐱,(cid:2919)x,𝐱(cid:1)(cid:2920)(cid:3439)==ϕφ((x𝐱(cid:2919)))(cid:2904)Tφϕ(cid:3435)(cid:0)𝐱x(cid:2920)(cid:3439)(cid:1) (1) (1) i j i j the LS-SVM classifier formulation results as: theLS-SVMclassifierformulationresultsas: (cid:2924) y(x)=sign(cid:34)(cid:3429)(cid:3533)α(cid:2919)y(cid:2919)K(x,x(cid:2919))(cid:3433) (cid:35) (2) n y(x) =sign ∑(cid:2869)αyK(x,x) (2) i i i The radial basis function (RBF) was selected a1s kernel function K (·,·): K(x,x)=exp[−‖x−x‖(cid:2870)⁄σ(cid:2870)] (3) Theradialbasisfu(cid:2919)nction(RBF)w(cid:2919)asselectedaskernelfunctionK(·,·): For the classification of the rhythms of the Physionet Challenge database, a multiclass model (cid:104) (cid:105) was adopted. The multiclass caKte(gxo,rxiz)at=ione xpprob−le(cid:107)mx −is xso(cid:107)l2v/edσ 2by a set of binary classifiers. We (3) i i adopted the “one-versus-one” coding [38], consisting in a set of binary classifiers, each discriminating Fboetrwtheeenc tlwasos icfilacsasteiso.n oftherhythmsofthePhysionetChallengedatabase,amulticlassmodelwas The LS-SVMlab toolbox [39] was used. First, the hyper-parameters (regularization γ and the RBF adopted. Themulticlasscategorizationproblemissolvedbyasetofbinaryclassifiers. Weadopted kernel dispersion σ2) were estimated and optimized by cross-validation (10-fold) selecting the values the“one-versus-one”coding[38],consistinginasetofbinaryclassifiers,eachdiscriminatingbetween which minimize the rate of misclassifications. Then, the parameters α and b of the SVM model were twoclasses. estimated. Overfitting of the dataset may occur and the estimated classifier may have poor The LS-SVMlab toolbox [39] was used. First, the hyper-parameters (regularization γ and the generalization capability. Therefore, we compared different LS-SVM classifiers using different RBFkerneldispersionσ2)wereestimatedandoptimizedbycross-validation(10-fold)selectingthe Electronics2018,7,199 7of14 values which minimize the rate of misclassifications. Then, the parameters α and b of the SVM modelwereestimated. Overfittingofthedatasetmayoccurandtheestimatedclassifiermayhave poorgeneralizationcapability. Therefore,wecompareddifferentLS-SVMclassifiersusingdifferent numbersoffeatures,pre-selectedbystep-wisediscriminantanalysis. Suchclassifiersweretrained onthe80%oftheavailabledatasetandthentheywerevalidatedontheremaining20%. Thefeature setleadingtothebestperformanceonthevalidationsetwasselected. Sincecross-validationisused tovalidatethehyper-parameterstotrainamodel,ratherthantovalidatethemodelitself,wethen selectedthebestparameterstore-traintheselectedmodelonalltheavailabledataset. Thisallowedus toreduceoverfittingusingahighernumberofcases. ThefinalperformanceofthealgorithmwasthenassessedbytheChallengeorganizersrunning thecodeonanindependenthidden“testset”,whichwasneverusedfortheclassifierestimation. FortheAFDB,abinaryclassificationwasappliedastherhythmswereonlycategorizedasNSRor AF.Inthiscase,theclassifierwastrainedonthe60%oftheAFDBandthenvalidatedontheremaining 40%. Inordertotestthegeneralizationcapabilitiesoftheapproach,thesamesetof30featuresselected forthePhysionetChallengedatabasewasusedforAFDB. 2.5. Scoring AccordingtothePhysionetChallenge2017scoringsystem[25],weevaluatedtheperformance ofthealgorithmusingtheF measure,whichistheharmonicmeanbetweensensitivityandpositive 1 predictivevalue. ThefollowingnotationproposedinCliffordetal.[25]isused: thetrueclassesfor Normal, AF, Other and Noisy rhythms are indicated by the uppercase letters “N”, “A”, “O” and “P”respectively, thepredictedclassesareindicatedbythelowercaseletters“n”, “a”, “o”and“p”. Thenotation“Xy”denotesthenumberofelementsoftheclass“X”assignedtotheclass“y”. TheF 1 valuewasobtainedforeachclassasfollows: 2×N NSR:F = n (4) 1n ∑N+∑n 2×A AFrhythms:F = a (5) 1a ∑A+∑a 2×O Otherrhythms:F = o (6) 1o ∑O+∑o 2×P p Noisyrhythms:F = (7) 1p ∑P+∑p Theglobalvaluewasthencalculatedasthemeanofthesinglevalues: F +F +F F = 1n 1a 1o (8) 1 3 Notably,theglobalscorewastheaverageofperformanceforNSR,AFrhythmsandotherrhythms withtheintenttoemphasizecorrectclassificationofusablerecordings,reducingtheeffectiveweight assignedtorecordingsthataretoonoisytobeanalyzed[25]. IntheAFDBonlyNSRandAFrhythms werepresentandsotheF wascomputedas: 1 F +F F = 1n 1a (9) 1 2 3. Results 3.1. PhysionetChallenge2017Database Theresultsreportedhereafterrefertothelatestresultsachievedonthefinalsetofground-truth labels(v3)ofthefollow-upphaseofthechallengeagreeduponbythechallengeorganizers,whichwere Electronics2018,7,199 8of14 madeavailableinNovember2017. Indeed,theseannotationswerecorrectedandrevisedcomparedto theofficialchallengeversionandsotheresultsaremorereliable. The stepwise discriminant analysis selected 30 discriminant features, which are described in Table1. Amongthese,theMawsdco(F=2114)andtheCosEn(F=1913)werethefeatureswiththe Electronics 2018, 7, x FOR PEER REVIEW 8 of 14 mostdiscriminantpower,accordingtothemulticlassANOVAanalysis. InIno rodredrert otou unnddeerrssttaanndd wwhhiicchh wweerree ththe emmosots stigsnigifnicifiancat nfetafteuarteus riens diinscdriimscirniamtiinnga teiancgh epaacihr opf air ofcclalassssees,s ,eevvereyr ybbetewtweeene nclcalsassesse sAANNOOVAVA wwasa salaslos opeprefrofromrmede. dM. aMwaswdcsod choadh atdhet hmeamxiamxailm Fa liFn in disdcirsicmriimnaintiantignNg SNRSRan adndA FArFh ryhtyhtmhms(sF (F= =1 01200280)8a) sassh sohwownnin inF iFgiugruere3 a3.aI.t Iits ips opsossisbilbeleto ton ontoictieceth tahtatth is feathtuisr efeiastuurseu aisl luyshuiagllhye hriignhAerF inrh AyFth rmhysthcomms pcoamrepdawreidth wNitShR N.SR. (a) (b) (c) (d) (e) (f) FigFuigruer3e. H3. isHtoigstroagmrasmofs Mofe aMneoafnth eofa bthsoel uatbeswoleuitgeh twedeigsuhctecdes ssiuvcecdesifsfievree ndcieffceorenntrcoel licnogntfroorllpinregm faotru re bepatrsemwaitthurceo mbepaetsn swaittohr ycopmaupseens(wateoirgyh ptaoupsteim (wizeaitgiohnt )o(pMtiamwizsadtcioon))( a()Maanwdscdoceof)fi (caie) natnodf csoaemffpicleieennt torof py (CsoaSmEpnl)e( ebn)tfroorpyN (aCnodSEAn)r (hby)t fhomr sN; hainsdto Ag rrahmytshomfsC; hoiSsEtongr(ca)masn odf MCoaSwEsnd (cco) a(dnd) fMoraOwsadncdo (Ad)r hfoyrt hOm s; hisatnodg rAam rhsyothfmroso; thmisetoagnrasmqus aorfe roofostu mcceeasns isvqeudarifef eorfe nsuceccse(sRsiMveS SdDiff)e(ree)nacensd (KRMatSzSFDr)a c(eta) laDndim Keantszi on (KFFrDa)ct(afl) DfoimrNenasniodn O(KrFhDy)t h(fm) fso.rN N: nanodrm Oa rl;hAyt:hamtrsi.a Nlfi: bnroirlmlatailo; nA;:O at:roiathl feirb.rillation; O: other. OnOeniem impoprotratnantta aimimo offt hthiiss ssttuuddyy wwaass ttoo aaccccuurraatteellyy ddisisccrirmimininataet eAAF Ffrformom otohtehr errhryhthymthsm ass athset he literature on this topic is still poor. According to the ANOVA, the CoSEn (F = 2563) and the Mawsdco literatureonthistopicisstillpoor. AccordingtotheANOVA,theCoSEn(F=2563)andtheMawsdco (F = 1933) were the features with the most discriminant power for the classification between these two classes as observed in Figure 3c,d. Both the two features resulted generally higher in AF rhythms compared with other rhythms. For the discrimination between normal and other rhythms, RMSSD results as the most powerful in the discrimination (F = 1712) followed by KFD (F = 1551) as shown in Figure 3e,f. Although there Electronics2018,7,199 9of14 (F=1933)werethefeatureswiththemostdiscriminantpowerfortheclassificationbetweenthese twoclassesasobservedinFigure3c,d. BoththetwofeaturesresultedgenerallyhigherinAFrhythms comparedwithotherrhythms. Forthediscriminationbetweennormalandotherrhythms,RMSSDresultsasthemostpowerful inthediscrimination(F=1712)followedbyKFD(F=1551)asshowninFigure3e,f. Althoughthereis awindowofoverlapbetweenthetwoclasses,otherrhythmshaveoftenahigherRMSSDcompared withNSR.Itshouldbenotedthat,despiteitsdiscriminantpowerindifferentiatingbetweennormal andotherrhythms,theRMSSDwasnotselectedbythestepwisediscriminantanalysis,asprobablyits discriminantinformationwasredundantconsideringtheotherincludedfeatures. Inallthefigures,theareasofthehistogramsforeachclassarenormalizedtoone,soareestimates oftheclass-conditionalprobabilitydensities. Usingthe30selectedfeatures,theLS-SVMclassifierprovidedthefollowingperformanceinthe multi-classclassificationonalltheavailabledatasetoftheChallenge(afterthere-trainofthemodel): F =0.93;F =0.85;F =0.83andF =0.87. FornoisyrecordingswegotF =0.83. Theconfusion 1n 1a 1o 1 1p matrixobtainedisreportedinTable2. Table2.ConfusionmatrixontheavailablesetofthePhysionetComputinginCardiologyChallenge 2017.N:normal,A:atrialfibrillation,O:otherrhythms,P:noisy.Lowercaselettersdenotetheclassifier predictedclasses.Thebackgroundgreencolorindicatescorrectclassifications. n a o p N 4923 16 129 8 A 23 635 95 5 O 459 74 1875 6 P 58 7 3 108 Onthehiddentestofthefollow-upphasetheperformancewas:F =0.919;F =0.824;F =0.746 1n 1a 1o andF =0.83. Withthisperformanceweobtainedthethirdplaceoutof84participantsinthefinalrank 1 ofthefollow-upchallenge(firstplace: F =0.921;F =0.857;F =0.766andF =0.848[40];second 1n 1a 1o 1 place: F =0.908;F =0.841;F =0.745andF =0.832[41](https://groups.google.com/forum/ 1n 1a 1o 1 #!topic/physionet-challenges/qA2iUfQmRtc). Acomparativesummaryofthealgorithms[40–43], whichobtainedthetopfivescoresinthefollow-upphaseofthePhysionetChallenge2017,isreported inTable3. Table3.Comparisonoftheresultsoftheproposedmethodswiththatoftheotheralgorithmsproposed forthePhysionetChallenge2017obtainedinthefollow-upphase. Reference NumberofFeatures ClassificationMethod OverallPerformance GradientBoostingclassifierand Teijeiroetal.[40] 42 RecurrentNeuralNetwork F1=0.848 Kropfetal.[41] 380 RandomForest F1=0.832 Billecietal.(proposedmethod) 30 LeastSquare-SupportVectorMachine F1=0.830 Dattaetal.[42] 150 Multi-layerCascadedBinary F1=0.8294 Plesinger[43] 60 NeuralNetwok+BaggedTreeEnsemble F1=0.8278 Notably, ourperformanceonthehiddentestsetoftheofficialchallengephasewas: F1n =0.911; F1a =0.784; F1o =0.739andF1 =0.812(obtainingthe12thplaceintheofficialrankinglist). Theincreaseinperformance achievedfromtheofficialphasetothefollow-upphasesuggeststheimportanceofhavingreliableannotationsfor thealgorithmtrainingandevaluation. 3.2. MIT-BHAFDatabase TheapplicationofthebinarySVMclassifierontheAFDBprovidedthefollowingperformance: F Normalrhythm=0.98; F AFrhythm: 0.99; GlobalF =0.98. Theconfusionmatrixobtainedis 1 1 1 reportedinTable4. Electronics2018,7,199 10of14 Table4.ConfusionmatrixontheAFDB.N:normal,A:atrialfibrillation;lowercaselettersdenotethe classifierpredictedclasses.Thebackgroundgreencolorindicatescorrectclassifications. n a N 5393 75 A 80 10,937 4. Discussion In this study, we proposed a novel approach for discriminating between normal, AF, other rhythms, and noisy ECG records. The algorithm achieved high global performance both in short-term(F =0.83)andlong-termrecordings(F =0.98).Importantly,theresultsofthisstudyshowed 1 1 the possibility of accurately detecting and discriminating AF in single short-lead ECG recordings obtainedwithasmartphone-connecteddevice. Inparticular,ascanbeobservedfromtheconfusion matrix,theclassifierseemsparticularlyefficientindiscriminatingbetweennormalAFrhythmsandAF fromotherrhythms,whileitislessefficientindiscriminatingnormalfromotherrhythms,asshownin Table2. Veryfewattemptshavebeenmade,outsidethePhysionetChallenge2017,todiscriminateAF fromtheotherarrhythmia. Thus,theclassificationofdifferenttypesofrhythmsisanimportantnovel aspectofthealgorithms,includingours,implementedforthechallenge. Comparedtotheotheralgorithmssubmittedforthechallenge,thestrengthofourapproachisthat itusedaquitesimplemethodforextractingtemporalandspectralfeaturesandapplyingtheLS-SVM forclassification. Thisallowedustoobtainagoodtrade-offbetweenaccuracyandcomputationaltime, thuspotentiallyallowingareal-timeapplicationofthealgorithm. Moreover,weusedalownumberof featurescomparedtotheotheralgorithms,asseeninTable3,whichfosteredagoodgeneralization capabilityofthealgorithm. Notably,theproposedalgorithmreliesonthecombinationofdifferenttypesoffeatures,obtained by RR analysis, PWA, and FSA for a better characterization and discrimination of AF. Moreover, other features extracted from the ECG or specifically from the QRS have been computed in order tobetterdiscriminateAFfromotherrhythms. Fewpreviousstudieshaveusedasimilarcombined approachforabinaryclassificationofAFandNSRobtaininghighperformance[14,15]. Theresults of our study suggest that a similar multidimensional approach can be extended to the multiclass problem. Indeed,accordingtotheresultsofacomparativestudyofAFdetectionalgorithms,theones thatcombineRRandatrialactivity(AA)analysisarethosewiththehighestpositivepredictivevalue, sowithbothahighspecificityandsensitivity[13]. Thisisrelevantwhentheaimisnotonlytodetect AFbutalsotodistinguishitfromotherrhythms. According to the ANOVA analysis, the features extracted from the RR series were the most discriminantones. Indeed,theAFdoesnotconveytotheventriclestherhythmofthesinusnode, butirregularstimulationduetotherecurrentpropagationofthedepolarizationwaveontheatria. Thus,theirregularityofAFmostlyaffectstheRR.Inaddition,theAAcanbeinfluencedbynoise[13] limitingthereliabilityoftheFSAfeatures. Inparticular,theMawsdcoandtheCoSEnwerethemostdiscriminantfeaturesfordistinguishing among the different classes. CoSEn also resulted as one of the two most discriminant features in classifyingNSRandAFrhythms. ThehighefficiencyofCoSEnindetectingAFconfirmsprevious studies[18]. Indeed,duringAF,electricaldischargesconductedfromtheatriumintotheventricles are irregular and disorganized so that randomness and entropy increase during an AF episode. Inparticular, ithasbeennoticedthatCoSEnisparticularlyrelevantwhenclassifyingshortlength recordings [32,44]. Importantly, we observed that CoSEn is also the most powerful feature in the discriminationbetweenAFandotherrhythms.Thefewstudieswhichattempttodiscriminatebetween AF and other type of rhythms showed the good discriminant power of this feature [22–24]. Thus, ourstudyconfirmstheimportanceofincludingthisfeaturetodiscriminateAFfromotherrhythms.

Description:
The authors achieved high performance on their own database of of algorithms to classify single short ECG lead recordings, as NSR, AF, other In this paper, we propose a novel approach for the classification of single-lead short ECG In Figure 1 the full process of the proposed method is reported.
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.