ECG-based heartbeat classification for arrhythmia detection

21 Pages·2016·2.76 MB·English

computer methods and programs in biomedicine 127 (2016) 144–164

journal homepage: www.intl.elsevierhealth.com/journals/cmpb

ECG-based heartbeat classification for arrhythmia

detection: A survey

Eduardo José da S. Luza, William Robson Schwartzb,
Guillerm o Cá ma ra- Cháve za, Davi d Meno ttia,c,∗

aUniversidadeFederaldeOuroPreto,ComputingDepartment,OuroPreto,MG,Brazil

bUniversidadeFederaldeMinasGerais,ComputerScienceDepartment,BeloHorizonte,MG,Brazil

cUniversidadeFederaldoParaná,DepartmentofInformatics,Curitiba,PR,Brazil

a r t i c l e i n f o a b s t r a c t

Articlehistory: Anelectrocardiogram(ECG)measurestheelectricactivityoftheheartandhasbeenwidely

Received27March2015 usedfordetectingheartdiseasesduetoitssimplicityandnon-invasivenature.Byanalyzing

Receivedinrevisedform theelectricalsignalofeachheartbeat,i.e.,thecombinationofactionimpulsewaveforms

8November2015 producedbydifferentspecializedcardiactissuesfoundintheheart,itispossibletodetect

Accepted17December2015 someofitsabnormalities.Inthelastdecades,severalworksweredevelopedtoproduce

automaticECG-basedheartbeatclassificationmethods.Inthiswork,wesurveythecurrent

Keywords: state-of-the-art methods of ECG-based automated abnormalities heartbeat classification

ECG-basedsignalprocessing bypresentingtheECGsignalpreprocessing,theheartbeatsegmentationtechniques,the

Heartbeatclassification feature description methods and the learning algorithms used. In addition, we describe

Preprocessing someofthedatabasesusedforevaluationofmethodsindicatedbyawell-knownstandard

Heartbeatsegmentation developedbytheAssociationfortheAdvancementofMedicalInstrumentation(AAMI)and

Featureextraction describedinANSI/AAMIEC57:1998/(R)2008(ANSI/AAMI,2008).Finally,wediscusslimitations

Learningalgorithms anddrawbacksofthemethodsintheliteraturepresentingconcludingremarksandfuture

challenges,andalsoweproposeanevaluationprocessworkflowtoguideauthorsinfuture

works.
©2015ElsevierIrelandLtd.Allrightsreserved.

formedbyasetofirregularheartbeats,hereincalledrhythmic
1. Introduction
arrhythmias.Theclassificationofnormalheartbeatsandthe

onescomposingtheformergroupareonthefocusofthissur-

Therearevarioustypesofarrhythmiasandeachtypeisassoci- vey.Theseheartbeatsproducealterationsinthemorphology

atedwithapattern,andassuch,itispossibletoidentifyand orwavefrequency,andallofthesealterationscanbeidentified

classify its type. The arrhythmias can be classified into two bytheECGexam.

major categories. The first category consists of arrhythmias Theprocessofidentifyingandclassifyingarrhythmiascan

formedbyasingleirregularheartbeat,hereincalledmorpho- be very troublesome for a human being because sometimes

logicalarrhythmia.Theothercategoryconsistsofarrhythmias itisnecessarytoanalyzeeachheartbeatoftheECGrecords,

∗ Correspondingauthorat:UniversidadeFederaldoParaná,DepartmentofInformatics,81.531-980Curitiba,PR,Brazil.

Tel.:[phone];fax:[phone].

E-mailaddresses:[email protected](E.J.d.S.Luz),[email protected](W.R.Schwartz),[email protected](G.Cámara-Chávez),

[email protected],[email protected](D.Menotti).

http://dx.doi.org/10.1016/j.cmpb.2015.12.008
0169-2607/©2015ElsevierIrelandLtd.Allrightsreserved.
computer methods and programs in biomedicine 127 (2016) 144–164 145

acquired by a holter monitor for instance, during hours, or learning algorithms found in literature for arrhythmia clas-

evendays.Inaddition,thereisthepossibilityofhumanerror sification. Section 7 presents the recommended evaluation

duringtheECGrecordsanalysis,duetofatigue.Analternative standard proposed by AAMI and describes the characteris-

istousecomputationaltechniquesforautomaticclassifica- ticsofthemostutilizeddatabases,indicatedbythestandard,

tion. toevaluatetheclassificationarrhythmiamethods.Section 8

Afullautomaticsystemforarrhythmiaclassificationfrom presentssomecommentsrelatedtotheissueofselectingdata

signals acquired by a ECG device can be divided in four for learning/evaluating models for arrhythmia classification

steps(seeFig.1),asfollows:(1)ECGsignalpreprocessing;(2) anditsimpactonthefinalresult.Finally,Section 9discusses

heartbeatsegmentation;(3)featureextraction;and(4)learn- thelimitationsandproblemsofthefieldandpointoutfuture

ing/classification.Ineachofthefoursteps,anactionistaken challengesfortheresearchcommunity.

andthefinalobjectiveisthediscrimination/identificationof

thetypeofheartbeat.

The first two steps of a such classification system (ECG
2. ECGsignal
signalpreprocessingandheartbeatsegmentation)havebeen

widely explored in the literature [2–6]. The techniques

employedduringthepreprocessingstepdirectlyinfluencethe Theheartisamusclethatcontractsinarhythmicalmanner,

final results, and therefore, should be carefully chosen. The pumpingbloodthroughoutthebody.Thiscontractionhasits

resultsrelatedtotheheartbeatsegmentationstep,inthecase beginningattheatrialsinenodethatactsasanaturalpace-

of QRS detection, are very close to optimal. However, there maker, and propagates through the rest of the muscle. This

is still room for exploration and improvements in the steps electricalsignalpropagationfollowsapattern[11].Asaresult

relatedtoclassification(featureextractionandlearningalgo- ofthisactivity,electricalcurrentsaregeneratedonthesurface

rithms).EventhoughtheproblemofECGdelineationisstill ofthebody,provokingvariationsintheelectricalpotentialof

open,itisnotsousefulforthemethodsintheliteraturesur- theskinsurface.Thesesignalscanbecapturedormeasured

veyedhere. withtheaidofelectrodesandappropriateequipment.

Thispaperpresentsasurveyofexistingstudiesfoundin The difference of electrical potential between the points

literature regarding the ECG-based arrhythmia classification marked by the electrodes on the skin, usually is enhanced

methodsanddiscussesthemaintechniquesusedforthecon- withtheaidofaninstrumentation(operational)amplifierwith

struction of these automatic systems as well as two main optic isolation. Then, the signal is submitted to a high-pass

paradigmsusedforevaluation:inter-patientandintra-patient filter; and as a second stage, submitted to an antialiasing

[7,8]. In addition, the most popular databases and the prob- low-pass filter. Finally, it appears in an analogical to digi-

lems related to the evaluation of current methods found in tal converter. The graphical registration of this acquisition

literaturearealsodiscussed.Fromthisdiscussion,aworkflow process is called electrocardiogram (ECG) (see Fig. 2). Since

isproposedtoguidetheevaluationprocessoffutureworks. AugustusDesiréWallerdemonstratedthefirsthumanECGin

Notethatthisworkflowforevaluationprocessconstitutesan 1887,theelectricalactivityofthehearthasbeenrecorded[12].

importantcontributionofthissurveywork.Intheliterature, Even so, the ability to recognize the normal cardiac rhythm

we find a survey of knowledge-based ECG interpretation [9] and/orarrhythmiasdidnotbecomeroutineinmedicalcheck-

reviewingmethodsproposedinthe20thcentury.Cliffordetal. upsuntil1960.

[1] performed an extensive survey on the methods used for Nowadays, there are many approaches to measure-

ECG signal analysis. Their study focused on the physiology ment/recordECG.daSilvaetal.[13]providedataxonomyof

ofthesignal,aswellasitsprocessingtechniques,mainlyon state-of-the-artECGmeasurementmethods:in-the-person,on-

thefeatureextractionandclassification.Inparticular,Clifford the-personandoff-the-person.

etal.[1]didnotfocusontheproblemofevaluatingmethods, Within the in-the-person category, there are equipments

whichisthedifferentialofourstudy,inadditiontoamoreup- designed to be used inside human body, such as surgically

to-dateliteraturereviewontheissue.Moreover,oursurveyon implantedones,subdermalapplicationsoreveningestedin

featureextractionbringsaspecialreviewonfeatureselection. theformofpills.Thesedevicesareusedwhenlessinvasive

Theremainingofthispaperisorganizedasfollows.Sec- approacharenotapplicable.

tion 2 introduces the fundamental aspects of ECG signals; Contrasting with the in-the-person category, there is off-

the state-of-art is described in Sections 3, 4, 5, and 6; and the-person category. Devices on this category are designed

theevaluationstandardsdevelopedbytheAssociationforthe to measure ECG without skin contact or with minimal skin

AdvancementofMedicalInstrumentation(AAMI)[10]andthe contact.Accordingto[13],thiscategoryisalignedwithfuture

databases recommended for these standards, together with trendsofmedicalapplicationwherepervasivecomputersys-

thecriticismsrelatedtothesystemsdevelopeduptodateand temsareareality.Examplesofsuchequipmentsaretheones

futurechallenges,arediscussedinSections 7,8,and9. basedoncapacitivedeviceswhichmeasuretheelectricfield

More specifically, Section 3 deals with the preprocess- changesinducedbythebodyallowingECGmeasurementat

ing techniques most utilized in ECG signals, while Section distanceof1cmormoreevenwithclothingbetweenthebody

4 presents the concept of segmenting heartbeats from the andthesensor[13–15].

ECG signals and its commonly employed techniques. Sec- ThemajorityofdevicesusedforECGmeasurementsarein

tion 5dealswiththekeypointforthesuccessofarrhythmia theon-the-personcategory.Devicesonthiscategorynormally

classification,i.e.,therepresentationofaheartbeatorthefea- requiretheuseofsomeelectrodesattachedtotheskinsur-

tureextractionprocess.Section 6discussesthemostpopular face.Examplesofsuchequipmentsarebedsidemonitorsand
146 computer methods and programs in biomedicine 127 (2016) 144–164

Fig.1– Adiagramofthearrhythmiaclassificationsystem.

Analog to
source Digit al ECG
Converter signal

(ADC)
Isolation High-Pass Band Filter Amplification Low-Pass Band Filter

Fig.2– SimplifieddisplayofthehardwareforthecaptureofECGsignals.Source:AdaptedfromRef.[1].

holters. Nowadays, the standard devices used for heartbeat

analysiscomefromthiscategory.

Onequipmentsbelongingintotheon-the-personcategory,

threeormoreelectrodesareusedtoobtainthesignal,inwhich

oneofthemservesasareferencefortheothers.Usually,the

referenceelectrodeisplacedneartherightleg.Assuch,there

canbedifferentvisionsoftheECGsignal,dependingonthe

pairofelectrodeschosentoconstructthesignal.Thesediffer-

entiatedvisionsaregiventhenameofleads.

Awidelyusedconfigurationofelectrodesisonecomposed

of5electrodes[16]:oneoftheelectrodesispositionedonthe

leftarm(LA),oneontherightarm(RA),oneontheleftleg(LL),

oneontherightleg(RL)andoneonthechest,totherightof

theexternal(VorV1).Anotherwidelyemployedsetupuses

10electrodes[16],where5extraelectrodes(besidesVorV1

on the chest and LA, RD, LL and RA on legs and arms) are

positionedonthechest(V2toV6)allowingaformationof12
Fig.3– Typical10electrodesconfiguration.
leads.The10electrodes(12leads)configurationcanbeseen

inFig.3.

From these configurations, several different leads can be
wave) and repolarization (T wave). The patterns provoked
constructed to visualize the ECG signal. For example, Fig. 4
byarrhythmiascandeeplychangethesewaves.Meanwhile,
illustrates3particularleads:(I)formedbytheelectricalpoten-
leadVanditscorrelateleads(V1,V2)favortheclassification
tialdifferencebetweentheLAandRAelectrodes;(II)formedby
ofventricularrelatedarrhythmias,sincethereareelectrodes
theelectricalpotentialdifferencebetweentheLLandRAelec-

trodes;and(III)formedbytheelectricalpotentialdifference

betweentheLLandLAelectrodes.

ThepreviouslydescribedleadIIisoneofthemostutilized

fordiagnosingheartdiseases.Ithighlightsvarioussegments

within the heartbeat, besides displaying three of the most

important waves: P, QRS and T (see Fig. 5). These waves

correspondtothefieldinducedbytheelectricalphenomena

occurring on the heart surface, denominated atrial depo-

larization (P wave), ventral depolarization (QRS complex Fig.4– MorphologyofthecurveforleadsI,IIandIII.
computer methods and programs in biomedicine 127 (2016) 144–164 147

QRS C omplex or60Hz),sincetheyallowquickandeasyapplicationofthe

R reject-band-filter.Theproblemwiththisapproachisthatthe

frequency of the noise is not always known, which can be

solvedbyapplyingfiltersforvariousfrequencybandstothe

T Peak T End signal. However, the indiscriminate use of filters, i.e., high-
Tpe Int erval pass and low-pass ones, distorts the morphology of the signal,
and many times, makes it unusable for diagnosing cardiac

diseases.Architectureswithadaptivefilters[23,24]werealso
ST
Segment employedfornoiseremovalfromtheECGsignals.However,

PR Interval according to Thakor and Zhu [25], this technique has con-

straintsanddoesnotoffergreatadvantagesovertheFIRdigital

P filters.Xueetal.[26]surmountsomeofthesedifficultiesby
U
using adaptive filters based on neural networks such that

thenoisereductionwassignificantlyimproved.Thisstrategy
Q Tasc. Tdesc.
proportionedbetterdetectionoftheQRScomplex,whencom-

S paredwiththesamemethodusinglinearlyadaptivefilters.

Inthelastdecade,manymethodsbasedonwavelettrans-
PR QT Interval forms ha ve been em ploye d to rem ove n ois e, since they
Segment
preserveECGsignalpropertiesavoidinglossofitsimportant

Fig.5– Fiducialpointsandvarioususualintervals(waves) physiological details and are simple from a computational

ofaheartbeat. point of view [27–29]. Sayadi and Shamsollahi [2] proposed

Source:Ref.[21]. a modification of the wavelet transform called the multi-

adaptivebionicwavelettransformanditwasappliedtoreduce

noiseandbaselinevariationoftheECGsignal.Thismethod

positioned on the chest, improving the registry of action presentedsuperiorresultswhencomparedtotheonesbased

potentialsonventricularmuscle. onthetraditionalwavelettransform.

Therefore,theleadsmostutilizedfortheautomaticheart- Othermethodshavealsopresentedinterestingresultson

beat and arrhythmia classification are leads II and V and noise attenuation. Sameni et al. [30] have proposed the use

themethodsthatuseacombinationofthesetwoleads(and of nonlinear Bayesian filters for ECG signal noise reduction,

othercombinations)aretheonesthatpresentthebestresults presentingpromisingresults.Anewalgorithmbasedonthe

to date [17]. In this sense, the recent work by Tomasic and ExtendedKalmanFilter[3],whichincorporatestheparame-

Trobec[18]reviewsmethodsworkingwithreducednumbers ters of the ECG dynamic model for ECG noise reduction and

ofleadsandapproachesforthesynthesisofleads,conclud- signalcompression,yieldedasignificantcontributionbecause

ing that the traditional 12 lead system can be synthesized the method showed the greatest effectiveness to date. Note

from a smaller number of measurements [19]. In contrast, thattheworksin[2,3,30]reporttheirresultsintermsofsignal

anotherstudypublishedbydeChazal[20]demonstratedthat tonoiseratio.

similareffectivenessforECGarrhythmiaclassificationcanbe Techniques for preprocessing the ECG signal are widely

obtainedatalessercomputationalcostwhenusingonlyone explored,butthechoiceofwhichmethodtouseisintrinsically

lead,comparedwithmethodsusingmultipleleads[7]. connected with the final objective of the research. Methods

Although on-the-person is the mainstream on devices focusing on the heartbeat segmentation from the ECG sig-

aiming heart diseases diagnoses, [13] have shown that data nal(i.e.,detectionoftheQRScomplex,otherwavesorfiducial

captured with off-the-person based devices can be highly pointsaimingatheartbeatdelimitation)tendtorequireapre-

correlated to data captured with traditional on-the-person processingthatisdifferentfromthemethodsfocusingonthe

based equipments. The authors claim that off-the-person automaticclassificationofarrhythmias.

basedequipmentscanextendpreventivemedicinepractices Table7sumarizesthemainreviewedreferencesofmeth-

byallowingECGmonitoringwithoutinterferenceondailyrou- odsaimingatheartbeatclassificationandthistableisexplored

tine. In that sense, we encourage researchers to build ECG further(Section 8).ThosemethodsfollowAAMIinstructions

databases based on off-the-person devices to evaluate and andthesameprotocoltoreporttheresults,butdifferentpre-

validateheartbeatclassificationmethodsforthatcategory. processingtechniquesareused.deChazaletal.[7]usedtwo

medianfilterstoremovebaselinewander.Onemedianfilter

of200-mswidthtoremoveQRScomplexesandP-wavesand
3. Preprocessing
otherof600mswidthtoremoveT-waves.Theresultingsignal

isthenfilteredagainwitha12-tap,low-passFIRfilterwith3-

Among all proposals for reducing noise in ECG signals, the dBpointat35Hz.Samepreprocessingisusedin[31–35,8,36].

simplestandmostwidelyusedistheimplementationofrecur- In[37]signalispreprocessedwith10thorderlowpassFIRfil-

sive digital filters of the finite impulse response (FIR) [22], ter. Ye et al. [38] used a wavelet-based approach to remove

whichwasmadecomputationallypossiblewiththeadvance baselinewander[39]andthenaband-passfilterat0.5–12Hz

in microcontrollers and microprocessors. These methods isappliedtomaximizeQRScomplexenergy.Bazietal.[40]pro-

workwellfortheattenuationoftheknownfrequencybands, posedtheuseofhighpassfilterfornoiseartifactsandanotch

suchasthenoisecomingfromtheelectricalnetwork(50Hz filterforpowernetworknoise.LinandYang[41]usesasecond
148 computer methods and programs in biomedicine 127 (2016) 144–164

orderlowpassfilterandtwomedianfilter.In[42],thesignalis Quad Level Vector [62], among others. Table 1 displays the

subtractedbyitsmeanandthennormalized.Escalona-Moran performance of some methods for heartbeat segmentation

etal.[43]usedtherawwavei.e.,nopreprocessingisapplied. that use the MIT-BIH database for evaluation. Note that the

NotethatthemethodscitedinTable7usedifferentprepro- SensitivitySEG (Se) and Positive predictivitySEG (+P) values do

cessingapproaches.However,theimpactoftheseapproaches notshowgreatdifferencesinthemethodsstudied.Itisimpor-

onautomaticarrhythmiaclassificationmethodsisnotclear. tant to highlight that the methods presented in this table,

The considered state-of-the-art methods do not even apply contemplatedalargespectrumofcomplexity,i.e.,fromvery

preprocessing on the signal. Although some studies exist simplemethodstomoreelaboratedones.

relatingpreprocessingtechniqueswiththefinalperformance Somealgorithmsalsoproposetoidentifyotherwavesasso-

of the automatic classification of arrhythmias, such as the ciated with heartbeats, such as the P wave and the T wave

work presented in [44], they are insufficient in number and [4,63–65], which can be useful for arrhythmia classification

more research in this area is encouraged. It is worth noting methods, since more information about the heartbeats can

that the state-of-the-art methods for automatic arrhythmia beobtained.

classificationdonotusestate-of-the-artpreprocessingmeth- Althoughheartbeatsegmentationisnotthemainfocusof

odssignaltonoiseratioimprovement. thissurvey,notethatthisstageisofparamountimportance

inthearrhythmiaheartbeatclassificationprocess,sincesome

errorsherearepropagatedtothefollowingstagesandhave
4. Segmentation
a strong impact in the final classification of the arrhythmia

system.However,alargemajorityofthereviewedresearches

Heartbeatsegmentationmethods(i.e.,detectionoftheRpeak hereinutilizeddatabasesinwhichtheeventsrelatedtoheart-

ortheQRScomplex)havebeenstudiedformorethanthree beatsegmentation,i.e.,thedetectionoftheRpeakortheQRS

decades [49,53,46,54,55] and the generations of these algo- complex,areidentifiedandpreviouslylabeled,reducingthe

rithms and newly developing methods reflect the evolution segmentation stage to a simple search of a labeled event in

of the processing power of computers. With the facility of thedatabase.Inthisway,theresultsreportedbytheseworks

usingfasterprocessingcomputers,authorsstoppedworrying disregard the impact of segmentation step even though the

about computational cost and started concentrating on the databaselabelingispronetohumanerrors.Therefore,eval-

heartbeatsegmentationaccuracy.Twomeasuresareusually uating the impact of different segmentation algorithms on

consideredforevaluatingtheaccuracyofheartbeatsegmen- automaticarrhythmiaclassificationmethodscanbeapromis-

tation:sensitivityandpositivepredictivity,whicharedefined ingresearchdirection.

as: Yeetal.[66]proposedatesttoinvestigatetherobustness

oftheirfeatureextractionmethodagainstonesegmentation
SensitivitySEG=TP/(TP+FN), (1) iss ue,th eR-pea kmislocate error.A Gaussia n-di stributedarti-

ficialjitterwasusedtoadderroronR-peakannotations.We

and suggest to other authors to incorporate such test in future

worksaimingautomaticheartbeatclassification.
PositivepredictivitySEG=TP/(TP+FP), (2)

5. Featureextraction
whereTP(TruePositive),FP(FalsePositive)andFN(FalseNeg-

ative)indicatethenumberofheartbeatscorrectlysegmented,
numb erofsegm en tationst ha tdonotcor respondt otheheart- Thefeatureextractionstageisthekeytothesuccessinthe
beats,an d numberofsegm enta tio nst hatwereno tp erfo rmed, hear tbeatcl assification ofth ea rrhy thm ia usin gtheEC G sig-
respec tive ly. nal. Any i nformation e xtr acte d from the heart beat use d to

Forafaircomparisonofthemethodsfocusingontheheart- discriminateitstypemaybeconsideredasafeature.Thefea-
beatse g men tation,asta nd ard database needsto be use d.The turescanbe ext racte dinvar iousforms dir ec tlyfrom the ECG
most utilized,andr e commend edbyAN SI/AA MI fo rthev ali- signa l’s m or phology in t he time doma in and/ or in the fre-
datio nofmed icale quipment[10], ist heMIT-BIH dat abas efor quency domain or f rom th e car diac rhyt hm. Mo st pop ular
arrhyth m iaanalys is[56]–int hisc as e,us edforhe artbeats eg- method spropos ed inlite ratur earedis cussedin Sect ion 5.1.
mentation, although oth e r d atab ases are a lso used, suc h as Event houghsom e worksreg ard featureext rac tionand fea-
thatofAHA [57]and thato fCSE[58]. How eve r,acco rding to turesele ctionas twoin tercha ngeab leterms ,thesetwo pro cess
Kohl er etal. [59], man yof the met hods presente dinthelit er- are i n fact di ffe rent . While feature extract ion is defi ned as
atured on ot use astan da rdiz eddataba se,oruse on lypa rtof the sta ge t hat involv es the descrip tion of a h ea rtbeat, fe a-
it,wh ich ma kesi t difficulttofair lycompar em eth ods. ture selec tion consists in c hoosing a su bs et with the m ost
Anap proach w idelyuse df orseg mentatio n,duetoitssim- repre sentative features wit htheobje ct ivetoim prov ethe clas-
plicity and prom ising result s, i s based on dig ital fil ter s for sification stage . Section 5.2 is d edicated to describe fea ture
the at tenua tion of th e noise an d rem ova l of the fluctu at- selection approa ches.

ingbaseline,nonlineartranslationsthatenhancetheRpeak

andadaptivedetectionthresholdwereproposedbyPanand 5.1. Featureextraction

Tompkins [49]. More sophisticated methods have also been

used,suchasmethodsbasedonneuralnetworks[53],genetic Themostcommonfeaturefoundintheliteratureiscalculated

algorithms[50],wavelettransform[60,61,4],filterbanks[46], fromthecardiacrhythm(orheartbeatinterval),alsoknownas
computer methods and programs in biomedicine 127 (2016) 144–164 149

Table1–Effectivenessofheartbeatsegmentationmethods.#and%standforabsoluteandpercentagenumbers.The

MIT-BIHArr.databaseisusedinallmethods.

Method Heartbeats TP FP FN Error Se +P

(#) (#) (#) (#) (%) (%) (%)

Martinezetal.[4] 109428 109208 153 220 0.34 99.80 99.86

MoodyandMark[45] 109428 107567 94 1861 1.79 98.30 99.91

Lietal.[5] 104182 104070 65 112 0.17 99.89 99.94

Afonsoetal.[46] 90909 90535 406 374 0.86 99.59 99.56

Bahouraetal.[6] 109809 109635 135 184 0.29 99.83 99.88

Leeetal.[47] 109481 109146 137 335 0.43 99.69 99.88

HamiltonandTompkins[48] 109267 108927 248 340 0.54 99.69 99.77

PanandTompikins[49] 109809 109532 507 227 0.71 99.75 99.54

Polietal.[50] 109963 109522 545 441 0.90 99.60 99.50

Moraesetal.[51] N/R N/R N/R N/R N/R 99.22 99.73

Hamilton[52] N/R N/R N/R N/R N/R 99.80 99.80

valuesfortheseintervals,consideringahealthyhumanbeing
Table2–TypicalfeatureofanormalECGsignal,witha
withnocardiacabnormalities.
cardiacfrequencyof60beatsperminute(bpm)ofa
healthy adult. Features extracted from the domain of time/frequency

together with the features of the RR interval appear as part
Feature Normalvalue Normalvariation
ofthemethodsthatproducedthehighestaccuraciesinlitera-
Pwave 110ms ±20ms
P Q/PRi nterval 160 ms ± 40 ms ture to date (see Table 7). The simplest way to extract features
QRSW idth 100 ms ± 20 ms in the time domain is to utilize the points of the segmented
QTi nterval 400 ms ± 40 ms ECG curve, i.e., the heartbeat, as features [73,74]. However, the
Am plitudeo fP 0.115m V ± 0.0 5mV use of samples of the curve as features is a technique that
Amplitude of Q RS 1.5 mV ± 0.5m V is n ot very effic ien t, si nce be sid es produ cin g a vector o f the
STlevel 0 mV ± 0.1 mV fe atur eswi thhighd imens ions(de pendingon t heamo un tof
Am plitu deofT 0.3 mV ± 0.2 mV samples used tore presenttheh eartbeat),i tsu ffer sfromse v-

Source:Ref.[1]. eralproblemsrelatedtothescaleordisplacementofthesignal

withrespecttothecentralpoint(peakR).

Aiming at reducing the dimension of the feature vector,

theRRinterval.TheRRintervalisthetimebetweentheRpeak various techniques have been applied directly on the sam-

ofaheartbeatwithrespecttoanotherheartbeat,whichcould ples that represent the heartbeat (in the neighborhood of

beitspredecessororsuccessor.Withexceptionofpatientsthat theRpeak)asprincipalcomponentanalysis(PCA)[75–77],or

utilizeapacemaker,thevariationsperceivedinthewidthof independentcomponentanalysis(ICA)[78–80],inwhichnew

theRRintervalarecorrelatedwiththevariationsinthemor- coefficientsareextractedtorepresenttheheartbeat.Chawla

phologyofthecurve,frequentlyprovokedbyarrhythmias[1]. [81]presentsacomparativestudybetweentheuseofPCAand

Thus,thefeaturesintheRRintervalhaveagreatcapacityto ICA to reduce the noise and artifacts of the ECG signal and

discriminatethetypesofheartbeatsandsomeauthorshave showedthatPCAisabettertechniquetoreducenoise,while

based their methods only on using the RR interval features ICAisbetteronetoextractfeatures.TheICAtechniqueenables

[67–69]. Variations of this feature are used to reduce noise statisticallyseparateindividualsourcesfromamixingsignal.

interferenceandareverycommon,e.g.,theaverageoftheRR TheECGisamixofseveralactionpotentialsandeachaction

intervalinapatientforacertaintimeinterval[70]. potential could be strongly related to an arrhythmia class.

LinandYang[41]haveshownthattheuseofanormalized The rationale behind ICA for ECG heartbeat classification is

RR-interval significantly improves the classification results. toseparatetheactionpotentialssourcesaswellasthenoise

OnlynormalizedRR-intervalsareusedinthatworkandthe sources.ThePCAtechniqueseparatesthesourcesaccording

resultsarecomparabletothestate-of-the-artmethodseven totheenergycontributiontothesignal.Thestudypresented

undertheinter-patientparadigm.Doquireetal.[71]confirmed in[81]suggestthatnoisesourcesonthisbasehavelowenergy

theefficiencyofnormalizedRR-intervalsbymeansoffeature and are difficult to isolate and that the individual sources

selectiontechniques. isolatedbyICAarepromisingfeaturesforECGclassification.

Otherfeaturesextractedfromtheheartbeatintervalsare Moreover, it has been shown that the combination of these

alsofoundinliterature,suchasotherdistancesbetweenthe twotechniques,i.e.,PCAfornoisereductionandICAforfea-

fiducialpointsofaheartbeat(herecalledECG-intervalsorECG tureextraction,canoffergreateradvantageswhencompared

segments), as can be seen in Fig. 5. Among these intervals, tousingonlyoneofthem.AnothertechniquebasedonPCA,

theQRSinterval,orthedurationoftheQRScomplex,isthe theKernelPrincipalComponentAnalisys(KPCA),wasusedby

mostutilized.Sometypesofarrhythmiasprovokevariations Kanaan et al. [82]. In that work, a comparison between PCA

intheQRSinterval,makingitagooddiscriminatingfeature andKPCAwasperformedanditwasconcludedthatKPCAis

[7,72].Itisworthmentioningthatthereexistotheralgorithms superiortothePCAtechniqueforclassifyingheartbeatsfrom

availabletodeterminethesefiducialpoints,suchastheone theECGsignal.AccordingtoKallasetal.[83],KPCAperforms

proposedbyLagunaetal.[63].Table2displaysthestandard better,duetoitsnonlinearstructure.
150 computer methods and programs in biomedicine 127 (2016) 144–164

themostpopularforECGsignalclassificationduetoitseasy

implementation.
BesidesDWT,continuouswavelettransform(CWT)hasalso

beenusedtoextractfeaturesfromtheECGsignals[99],since

itovercomessomeoftheDWTdrawbacks,suchasthecoarse-

ness of the representation and instability. However, CWT is

notlargelyusedduetothefactthatitsimplementationand

its inverse are not available in standard toolboxes (such as

MATLABwaveletToolbox)andCWTshouldbecarefullydis-

cretized for the use as a CWT analyzer. In addition, even

thoughAddison[99]emphasizesthehighcomputationalcost

asadisadvantageforusingCWT,ithasbeenemployedsuc-

cessfully even on simple medical equipments for at least a

decade.Finally,Addison[99]defendstheuseofDWT,together
Fig.6– Featurenumberreductionbymeansof
with CWT, because they offer gain over the methodologies
interpolation.
used nowadays, in which the authors use only one of the
Source:Ref.[7].
transforms.
AccordingtoGülerandÜbeyli[44],thechoiceofthemother

wavelet function used for feature extraction is crucial for

Özbay et al. [84] used clustering techniques directly in thefinalperformanceoftheclassificationmodel.Thischoice

points sampled from the curve to reduce from 106 samples should be carefully analyzed in order not to lose important

to67clusters/points.Theauthorsalsousedaclusteringtech- ECGsignaldetails.Besidesthechoiceofthemotherwavelet

niquetoincreasethenumberoffeaturesto212,buttherewere function, the order of filter and level of decomposition are

nosignificantdifferencesintheresults. parametersthatinfluencethefinalresultsofthearrhythmia

Asletal.[85]usedGeneralizedDiscriminantAnalysis(GDA) classification.Daamoucheetal.[100]proposedtheuseofthe

toreducethedimensionsofthefeaturesoftheheartbeatinter- ParticleSwarmOptimization(PSO)techniqueforoptimizing

valtypetoclassifyrhythmicarrhythmias.Theyreportedan these parameters, and concluded that this process improve

accuracycloseto100%forthistypeofarrhythmiausingthe thefinalresults.

MIT-BIHdatabase.However,theauthorsdidnottakecareto Intheliterature,variousstatisticalfeaturesextractedfrom

separatetheheartbeatsofthesamepatientusedduringtrain- the coefficients of wavelet transform are proposed, such as

ing and testing (intra-patient paradigm), which is a serious mean, standard deviation, energy [44] and coefficient vari-

concerndiscussedfurther.Theinter-patientparadigmshould ance[101].Thesefeatureshaveagreatadvantagesincethey

beconsideredforamorerealisticscenario. areimmunetothevariationsoffiducialpointmarking.Some

Simplertechniques,suchasinterpolation,havealsobeen authorsusedtechniquestoreducethespaceofthefeatures

usedtoreducethenumberofpointsrepresentingtheheart- afterapplyingthewavelettransform,suchasintheworkof

beat.AnexampleofthistechniqueispresentedbydeChazal Songetal.[102]whocomparedthePCAandlineardiscrim-

et al. [7], in which the heartbeat, originally represented by inant analysis (LDA) techniques for dimensional reduction

250samples(approximately600msofthecurve,sampledat after the use of wavelet transform. Wang et al. [103] and

360Hz), was sub-divided and presented in 18 samples (see Polat & Günes¸ [104] also employed PCA to reduce features

Fig. 6). In the literature, the sub-sampled ECG wave is also formedbywaveletcoefficientsandalsoreportedasignificant

calledmorphologyormorphologicalfeatures. improvementtheirresults.AccordingtoGülerandÜbeyli[44],

Recently,randomprojectionshavealsobeenemployedfor the Daubechies wavelets are the most appropriated mother

such aim, as in [86,42]. Huang et al. [42] show that features wavelets for ECG heartbeat classification. Among them, the

extractedwithrandomprojectionsproducedresultscompa- Daubechiesoforder2offersthebestaccuracy.

rabletothestate-of-the-artmethods,evenwhenconsidering Althoughmanytechniqueshavebeenproposedtoextract

theinter-patientparadigm. andreducefeaturesfromECGsignalsaimingheartbeatclassi-

Othertechniqueshavealsobeenemployed,suchaslinear fication,onlyafewofthemhaveconsideredtheinter-patient

predictive coding [87], high order accumulates [88,89], clus- paradigmasonecanseeinTable7.Therefore,itisdifficultto

tering[84,90,91],correlationdimensionandlargestLyapunov evaluatewhetherfeaturesextractedwithPCA,ICA,GDAand

exponent[92,93],Hermitetransform[94],localfractaldimen- othersareusefultodiscriminatepatientsorheartbeats.

sion[95]. Thevarianceoftheautocorrelationfunctionisconsidered

Althoughvarioustechniqueshavebeenconsidered,most tobeameasureofsimilarityorcoherencebetweenasignal

of the studies presented in literature use wavelet trans- and its shifted version [101]. This technique is used for fea-

formsandresearchersclaimthatthisisthebestmethodfor ture extraction from wavelet coefficients [101,37], and have

extractingfeaturesfromtheECGsignal[44,96,97].Thewavelet demonstratedtobeeffectiveinthediscriminationofarrhyth-

transformallowsinformationextractionfrombothfrequency micheartbeats.

and time domains, different from what is usually achieved Thevectorcardiogram(VCG)isarepresentationoftheECG

by the traditional Fourier transform [98] which permits the signalintwodimensionsthatintegratesinformationfromtwo

analysis of only the frequency domain. Within the types of leads (see Fig. 7). Features extracted with VCG were used in

wavelet transform, the discrete wavelet transform (DWT) is [105,31,106,37].AccordingtoGoldbergeretal.[107],heartbeat
computer methods and programs in biomedicine 127 (2016) 144–164 151

Maretal.[34]workwasonthemaintenanceofaccuracywith

theuseofareducednumberoffeatures.

Feature selection techniques can bring various benefits

for the classification methods, such as the increase of the

generalizationpoweroftheclassificationalgorithmsandthe

reductionofthecomputationalcost,duetothefactthatthey

useasmallernumberoffeaturestoconstructthefinalmodel

[34].However,intheworksanalyzedinthissurvey,thesetech-

niqueswerelittleexplored.

Doquireetal.[71]comparewrapperfeatureselectiontech-

nique against a filter feature selection technique and more

than 200 types features (dimensions) are considered for the

task.Thewrapperfeatureselectionisusedwiththeweighted

LDmodelusingaforward-backwardsearchstrategy.Thefilter

technique employed is the mutual information in conjunc-

tionwithrankingapproachandweightedSVM(SupportVector

Machines).Accordingtotheauthors,resultshaveshownthat

higherfiguresareobtainedwhenaverysmallnumberoffea-

Fig.7– VCGsetupusingtwoheartbeatsofMIT-BIH’srecord tures are selected. They stressed that the most important

202. featuresappearsareR-Rintervals,theamplitudeandlength

Source:AdaptedfromRef.[37]. oftheTwave,and2nd-orderstatistics.Alsotheyclaimedthat

themutualinformationcriterionisapowerfultoolforfeature

selectioninthisscenario.

classificationcategorizedasSupraventricularectopicbeat(SVEB) According to Zhang et al. [35], many features are associ-

andVentricularectopicbeat(VEB)(arrhythmicheartbeats)can atedwithmathematicalinterpretationanddonothaveaclear

be favored by information from leads of type V1, V2 or V4. meaning to physicians. Usually, the authors employ several

Because of this, it is believed that the features extracted by combined features and the understanding of which feature

VCG(combinedwithleadsIIandV1)canhelptobetterdis- contributestodetectionofwhichclassofheartbeatisalsonot

criminateminorityandimportantarrhythmicclassessuchas clearintheliterature.Aimingthat,Zhangetal.[35]proposed

SVEBandVEB. a heartbeat class-specific feature selection scheme to allow

the investigation of feature contribution for each arrhyth-

5.2. Featureselection mia/heartbeat class. Thus, we suggest the incorporation of

this approach on works aiming heartbeat classification. It

AccordingtoLlamedoandMartinez[37],manyauthorshave couldbringimportantcontributiontotheliteraturebyallow-

used techniques that reduce the feature space, but few ingbetterunderstandingofcorrelationamongheartdiseases

have investigated techniques for feature selection in the andfeaturesextractedfromECG.

context of arrhythmia classification. Llamedo and Martinez State-of-art techniques for attribute selection, such as

[37] employed, for the first time in literature, an algorithm Genetic Algorithms (GA) [109,110] and particle swarm opti-

for feature selection by using floating sequential search for mization (PSO) [111,112] can also provide promising results

arrhythmia classification. This method interchanges algo- andshouldbebetterinvestigatedinfutureworks.

rithms executing forward and backward searches to obtain

a set with the most robust features and avoid local optima
6. Learningalgorithms
in the feature space. The proposed method achieved better

resultsthanthestate-of-artmethodusingonlyeightselected

features. Oncethesetoffeatureshasbeendefinedfromtheheartbeats,

Recently,Maretal.[34]alsoperformedfeatureselectionby modelscanbebuiltfromthesedatausingartificialintelligence

using the floating sequential search [108]. In that study, the algorithmsfrommachinelearninganddataminingdomains

authorsanalyzedasetofpossibilitiesofthefeatureselection, [113–115]forarrhythmiaheartbeatclassification.

searchingforatrade-offbetweenthenumberoffeaturesand Thefourmostpopularalgorithmsemployedforthistask

accuracy. The aim of that research was to make a specially andfoundintheliteratureare:supportvectormachines(SVM)

developedmethodadequateforambulatorymonitoring;that [40,38,66], artificial neural networks (ANN) [34,116,69] and

is,tobespeciallyusefulinrealworldapplications.Forsuch linear discriminant (LD) [7,37,17], and Reservoir Computing

aim, an objective function optimized by a feature selection With Logistic Regression (RC) [43]. Note that the state-

method, was especially developed to be an indicator of the of-the-art method aiming heartbeat classification uses RC

qualityofthearrhythmiaclassificationsfromanECGsignal. algorithm.

In addition to the linear discriminant (LD) classifier used in Due to their importance for cardiac arrhythmic classifi-

previousworks,Maretal.[34]employedamulti-layerpercep- cation, these four classifiers (SVM, ANN, LD, and RC) are

tron.However,neitheroftheseresultswerebetterthanthose discussed in the next subsections (Sections 6.1, 6.2, 6.3 and

proposedbydeChazaletal.[7]andtheworkofLlamedoand 6.4).Then,Section6.5reviewsothertechniquesthatalsohave

Martinez[37]intermsofaccuracy.Nonetheless,thefocusof beenemployedtoarrhythmiaclassification.
152 computer methods and programs in biomedicine 127 (2016) 144–164

6.1. Supportvectormachines(SVM) ofclassifiersnotonlyreducestheoverallerrorintheneural

networks,butalsoreducestheincidenceoffalsenegatives.

SVM is one of the most popular classifiers found in litera-

ture for ECG-based arrhythmia classification methods. Park 6.3. Lineardiscriminants(LD)

et al. [33] used SVM and validated the method according

to AAMI standards and the data set split scheme proposed TheLinearDiscriminantisastatisticmethodbasedonthedis-

by de Chazal et al. [7]. These same authors used SVM in a criminantfunctions[114].Suchfunctionsareestimatedfrom

mock-hierarchyconfigurationtoresolvetheimbalanceofthe atrainingsetofdataandtrytolinearlyseparatethefeature

MIT-BIHdatabase,andreportedpromisingvalues.deLannoy vector, being adjusted by the weight vector and a bias. The

etal.[32]managedtoovercometheimbalanceoftheMIT-BIH criteria for calculating the weight vector varies according to

databasewithSVM,alternatingtheobjectivefunctionforeach the model adopted. In [7], the parameters were determined

class(WeightedSVM).Expressivegainswerereportedforthe usingthemaximum-likelihoodcalculatedfromtrainingdata.

SVEBandFclasses. Lineardiscriminantsaretheclassifiersmoreusedinmeth-

Various approaches with SVM variations have been pro- odsthatfollowtheschemeproposedbydeChazaletal.[7]and

posed,suchasacombinationofthefuzzytheorytorefineSVM recommended by AAMI. The authors of that research claim

classification[117],combinedwithanensembleofclassifiers thattheclassifierwaschosenforitssimplicityandforthefact

[42],geneticalgorithmscombinedwithrestrictedfuzzySVM thattheydidnotwanttoemphasistheclassifier,butinstead,

[118]andleastsquaresSVM[104].Huangetal.[42]usedthe theproposedfeatures.Amongstitsadvantages,LDcaneasily

SVMinahierarchicalmannerwithamaximumvotingstrategy overcomeproblemsgeneratedbytheimbalanceofthetrain-

andreportsignificantlyimprovements. ingset(adifficultypresentedbyapproachesbasedonSVM).

MoavenianandKhorrami[119]proposedtheuseofanew Whenusingtheschemeproposedin[7],itisagreatchallenge

kernelfunctionforcapturingdatafromSVM.Inthatwork,it totuneSVMandMLPclassifierstoobtainpromisingclassifica-

was used the same methodology for comparing the results tioneffectivenessfortheminoritySVEBandVEBclasses(see

obtained from a SVM and a Multilayer Perceptron Artificial Table9).Moreover,theLDclassifierrequireslesstrainingtime,

NeuralNetwork(MLP-ANN).WhileSVMwasmoreefficientin ifcomparedtoSVMandMLP,asitisnotiterative.Thatis,it

executiontime,bothinthetrainingandinthetesting,MLP simplycalculatesstatisticsfromthetrainingdataandthen,

performedbetterintermsofaccuracy,Sensitivity(Se),positive theclassificationmodelisdefined.

prediction(+P)andfalsepositiverate(FPR).

Since SVM presents a negative behavior for imbalanced 6.4. Reservoircomputingwithlogisticregression(RC)

classes,databasebalancingtechniquesforthetrainingphase,

whicharelittleexploredforthisproblem,canbestudiedin AccordingtoRodanandTinˇo[123],reservoircomputingmod-

futureresearch,asforexample,moresophisticatedsampling els are dynamical models aiming to process a time series

techniques,i.e.,SyntheticMinorityOver-samplingTechnique signal in two parts: represent the signal through a non-

(SMOTE)[120]. adaptabledynamicreservoirandadynamicreadoutfromthe

reservoir.MoredetailsregardingRCcanbefoundin[124].

The state-of-the-art method for heartbeat classification

6.2. Artificialneuralnetworks(ANN) uses RC [43]. According to Escalona-Moran et al. [43], their

approachusesasimplenonlineardynamicalelementsubject

TheANNarchitecturesmostlyusedforarrhythmiaclassifica- toadelayedfeedbackwhereeachpointoftheECGsignalis

tionareMultilayerPerceptrons(MLP)andProbabilisticNeural sampledandheldduringonedelaytimeandthenmultiplied

Networks (PNN). According to Yu and Chen [101], models by a binary random mask. The learning process is accom-

constructedwithPNNarecomputationallymorerobustand plished with logistic regression. The technique appears to

efficientthanthetraditionalMLP.However,in[121,84,122,88], be robust to the class imbalance of the dataset. Besides, it

it was proposed a hybrid neuro-fuzzy network methods in achievesthebestresultsintheliteraturetodate(seeTable7).

order to minimize the problems of MLP, increasing its gen- Inaddition,theauthorsalsoclaimthatthetechniqueissuit-

eralizationandreducingitstrainingtime. abletoimplementinhardwareduetoitslowcomputational

ManyotherapproachesbasedonANNhavebeenproposed. cost,whichallowsthedevelopmentofrealtimeapplications

Güler and Übeyli [44] used combined neural networks in forheartbeatclassification.

ordertoobtainamoregenericmethodfromamoresophis-

ticated form of cross-validation. However, of all the articles 6.5. Othertechniques

mentioned in this study, only that of Mar et al. [34] used

MLP with a more fair evaluation protocol by applying the Manyothermethodsforarrhythmiaclassificationhavebeen

patient division scheme proposed by de Chazal et al. [7]. developed using other machine learning and data mining

Thusbyusingthereportedresultsintheworksofthemeth- algorithms,suchasdecisiontrees[125,126,68],nearestneigh-

ods that utilizes ANN as classifier is impossible to makes a bors[127–129],clustering[73,130,131],hiddenMarkovmodels

fair comparison. Finally, Mar et al. [34] compared MLP with [132,133], hyperbox classifiers [105], optimum-path forest

Linear Discriminants and found that MLP was significantly [134], conditional random fields [8] and rules-based models

superior. [135,67,136].
Combiningclassifiershadbeenlittleexploredforthetask Algorithms with a lazy approach, such as the k Near-

in question. According to Osowski et al. [91], a combination est Neighbors (kNN), are not much used for the problem of
computer methods and programs in biomedicine 127 (2016) 144–164 153

arrhythmiaclassification,sincetheirefficiencyisintimately effectiveness,whencomparedtoothermethodsproposedin

connectedtopreviousknowledgetoperformtheclassification literature.However,notestusingafairercomparisonscheme,

ofeachsamplethatisrepresentedbythecompletetraining suchastheoneproposedbydeChazaletal.[7],andtherec-

set,whichleadstoahighcomputationalcostduringthetest- ommendationsofAAMI,wasdonewithmethodsthatusea

ingphase.Thiscostcaninvalidateitsusefordiagnosisinreal setofrules.ThissubjectisdiscussedindepthinSection 8.

time.MishraandRaghav[95]usedaclassifierbasedonkNN Using a few discriminative features from previous works

and reported promising results, however the computational [71,37],deLannoyetal.[8]proposedtheuseofweightedCondi-

costwasnotmentioned.Inotherworks,alsobasedonkNN, tionalRandomFieldsfortheclassificationofarrhythmiasand

intheliterature[127,92,128,137,129],noonepresentedamore comparedtheapproachwithSVMsandLDs.Theexperiments

fairevaluationprotocolforcomparisonofmethodsastheone demonstrated that the proposed method obtains promising

proposedbydeChazaletal.[7],andnoonealsofollowedthe results for the minority arrhythmical classes (SVEB e VEB).

AAMIrecommendations.Inaddition,thecomputationalcost However, the relatively low efficiency for the normal class

ofthesemethodswasnotinvestigated. (80%) represents a problem when used in real life scenario

Clustering techniques are widely used along with Artifi- (inter-patientparadigm),sincemanyhealthyheartbeatswill

cialNeuralNetworks.AccordingtoÖzbayetal.[84],theycan beclassifiedasarrhythmic.

improve the generalization capacity of the neural networks

and diminish the learning time. Some works used unsu-
7. DatabasesandtheAAMIstandard
pervised clustering techniques to agglomerate all of the

heartbeatsintherecordofagivenpatientintoclusters[131]
Variousdatabasesarecomposedofcardiacheartbeatgrouped
and the final classification of each cluster, i.e., the heart-
inpatientsrecordsfreelyavailablethatpermitsthecreationof
beats of that group, is then defined by a human specialist
astandardizationfortheevaluationofautomaticarrhythmia
[138,130,73].Otherworksinthissameway[139,140,17],using
classification methods. This standardization was developed
lineardiscriminantasclassifiersandfairevaluationschemes
byAAMIandisspecifiedinANSI/AAMIEC57:1998/(R)2008[10]
present promising results which are reliable for real-world
anddefinedtheprotocoltoperformtheevaluationstomake
applications.Itisimportanttonotethatthissemi-automatic
suretheexperimentsarereproducibleandcomparable.
(or patient-specific paradigm) and promising approach still
Theuseoffivedatabasesisrecommendedbythestandard-
dependsonahumanspecialist.
ization:
HMM is widely used to audio and speech signal anaysis

andrecognition[141,142].Coastetal.[132]usedHMMforthe
arrh ythmia clas sification proble m , o ther s tudie s hav e u sed • MIT-BIH:TheMassachusettsInstituteofTechnology–BethIsrael

thistechniquetoanalyzeECGsignals.Forinstance,Andreao HospitalArrhythmiaDatabase(48recordsof30mineach);
etal .[143]valid at edtheu seof HMMfo rEC Ganalysis inmed- • EDB: Th e European Society o f Ca rdiology ST -T Dat abase (90

icalclinics(realworld). recordsof2heach);
T he Op timu m-path Forest (OPF) classifier was used for • AHA:Th eA m e ricanHeartAssociationDatabaseforEvaluationof

arrhythmiaclassificationforthefirsttimebyLuzetal.[134]. VentricularArrhythmiaDetectors(80recordsof35mineach);
In that wor k, the OPF p erfo rma nce, in te rm s o f c om puta- • CU:TheCr eightonUniv ersitySus tain edVentr icu lar Arrh ythmia

tionalcostandoverallaccuracy,wascomparedtootherthree Database(35recordsof8mineach);
classifi ers: Baye sian,S VMandM LP.E xperiment ss howed that • NST: The No ise Stres s T es t Da tabase (12 records of ECG of

OPFobtained,inaverage,comparableresults,revealingitasa 30mineach,plus3recordswithnoiseexcess);

promissoryapproach.

Methodsthatuseadecisiontreeallowaninterpretationof The most representative database for arrhythmia is the

thedecisionsmadebythemodel[68].However,thistypeof MIT-BIH,andbecauseofthis,ithasbeenusedformostofthe

methodisnotefficientforcontinuousfeatures(belongingto publishedresearch.Itwasalsothefirstdatabaseavailablefor

a set of real numbers) [144,145] and feature vectors of large thisgoalandhasbeenconstantlyrefinedalongtheyears[148].

dimensions[146].Thus,methodsthatusedecisiontreescon- Themajorityoftheheartbeatsrecordedinthesedatabases

sideronlyafewfeatures.Forexample,in[68],onlythefeatures haveannotationsassociatedwiththetypeofheartbeatorthe

intheRRintervalwereusedinthedecisiontree.Meanwhile events. These heartbeat annotations, as much for the class

thehyperboxclassifiers,besidesprovidinghighlevelofinter- andforthefiducialpoints(e.g.,pointR,maximumamplitude

pretation of the classification rules, are also more efficient of the heartbeat) are fundamental for the development and

for higher dimension feature vectors [105]. Mert et al. [147] evaluationofautomaticarrhythmiaclassificationmethods.

used a combination technique of bagging and decision tree. TheANSI/AAMIEC57:1998/(R)2008standardalsospecifies

According to the authors, the Bagged Decision Tree demon- howannotationsshouldbedoneinthedatabases.Anexample

stratedgreateraccuracyandabettercapacitytodiscriminate canbeseeninFig.8,inwhichthereistheleadIIattheupper

theclasses. partofthefigure,leadV1atthelower,andsomeannotations

The methods with the greatest interpretation level are inthecenter.Noteworthyisthefactthatitisrecommended

the ones that use a set of rules. The set of rules presented thatrecordsofpatientsusingpacemakersshouldnotbecon-

by Tsipouras et al. [135,67,136] was obtained together with sidered.Inthisdatabase,4patients/recordshavethisproperty

cardiologists and are related to a morphological tachogram anditsrespectiveheartbeatsshouldberemoved.Inaddition,

for arrhythmic events. Methods constructed in conjunction segmentsofdatacontainingventricularflutterorfibrillation

withrulesusuallypresentaworserperformance,intermsof (VF)shouldalsobeexcludedfromtheanalysis.