Manuscript_OyildirimR2Clear.docx

Manuscript_OyildirimR2Clear.docx

《Manuscript_OyildirimR2Clear.docx》由会员分享，可在线阅读，更多相关《Manuscript_OyildirimR2Clear.docx（31页珍藏版）》请在冰豆网上搜索。

Seediscussions,stats,andauthorprofilesforthispublicationat:

AnovelwaveletsequencebasedondeepbidirectionalLSTMnetworkmodelforECGsignalclassification

ArticleinComputersinBiologyandMedicine·March2018

DOI:

10.1016/pbiomed.2018.03.016

CITATIONS

118

READS

5,079

Someoftheauthorsofthispublicationarealsoworkingontheserelatedprojects:

SmartGridViewproject

PowerqualitymonitoringsystemsViewproject

AllcontentfollowingthispagewasuploadedbyÖzalyıldırımon22July2018.

Theuserhasrequestedenhancementofthedownloadedfile.

ANOVELWAVELETSEQUENCESBASEDONDEEPBIDIRECTIONALLSTMNETWORKMODELFORECGSIGNALCLASSIFICATION

ÖzalYILDIRIM（*correspondingauthor）

1ComputerEngineeringDepartment,EngineeringFaculty,MunzurUniversity,Tunceli,Turkey

*Correspondingauthor.Tel.:

+90-428-2131794;fax:

+90-428-2131861.

E-mailaddress:

oyildirim@munzur.edu.tr,yildirimozal@

29

ANovelWaveletSequencesBasedonDeepBidirectionalLSTMNetworkModelforECGSignalClassification

ÖzalYILDIRIM

ComputerEngineeringDepartment,MunzurUniversity,Tunceli,Turkey

oyildirim@munzur.edu.tr,yildirimozal@

Abstract

Long-shorttermmemorynetworks（LSTMs）,whichhaverecentlyemergedinsequentialdataanalysis,arethemostwidelyusedtypeofrecurrentneuralnetworks（RNNs）architecture.Progressonthetopicofdeeplearningincludessuccessfuladaptationsofdeepversionsofthesearchitectures.Inthisstudy,anewmodelfordeepbidirectionalLSTMnetwork-basedwaveletsequencescalledDBLSTM-WSwasproposedforclassifyingelectrocardiogram（ECG）signals.Forthispurpose,anewwavelet-basedlayerisimplementedtogenerateECGsignalsequences.TheECGsignalsweredecomposedintofrequencysub-bandsatdifferentscalesinthislayer.Thesesub-bandsareusedassequencesfortheinputofLSTMnetworks.Newnetworkmodelsthatincludeunidirectional（ULSTM）andbidirectional（BLSTM）structuresaredesignedforperformancecomparisons.ExperimentalstudieshavebeenperformedforfivedifferenttypesofheartbeatsobtainedfromtheMIT-BIHarrhythmiadatabase.ThesefivetypesareNormalSinusRhythm（NSR）,VentricularPrematureContraction（VPC）,PacedBeat（PB）,LeftBundleBranchBlock（LBBB）,andRightBundleBranchBlock（RBBB）.TheresultsshowthattheDBLSTM-WSmodelgivesahighrecognitionperformanceof99.39%.Ithasbeenobservedthatthewavelet-basedlayerproposedinthestudysignificantlyimprovestherecognitionperformanceofconventionalnetworks.Thisproposednetworkstructureisanimportantapproachthatcanbeappliedtothesimilarsignalprocessingproblems.

Keywords:

Long-ShortTermMemory,RecurrentNeuralNetworks,DeepLearning,ECGsignals.

1.Introduction

Improvementsinthefieldofmachinelearninghaveallowedforthedevelopmentofcomputer-basedintelligentdecisionsupportsystemsandtheirimplementationmoreefficientlyinmanyareas.Thedevelopmentofintelligentsystemsinthehealthfieldisattractiveintermsoftheamountofdataandtheimportanceofthedataitcontains.Researchersindifferentdisciplines,specificallyintermsofapplications,frequentlystudybiomedicaldata,includingimageandsignaldata.Electrocardiogram（ECG）signalsareoneofthemostfrequentlystudieddata.ECGsignalsincludeinformationaboutheartfunctionandheartconditions.Therefore,monitoringandrecognitionofECGsignalsisaremarkabletopicinthebiomedicalfield[1].

StudiesperformedonECGsignalsaregenerallyexaminedintwoparts:

detectionandclassification.StudiesondetectionconcentrateontheproblemofdeterminingheartbeatwithintheECGdataobtainedforacertainperiodoftime.Thesedetectedbeatsaremarkedintheraw

dataforanalysis.Itisanareaparticularlyemphasizedforthedevelopmentofwearablebiomedicalmonitoringsystems[2,3].Threshold-basedmethods[4],digitalfilter-basedmethods[5,6],andwavelettransform（WT）[7-10]havebeenappliedinheartbeatdetectionscenarios.

TheclassificationofdetectedECGsignalsisanotherimportantstep.Atthisstage,automaticidentificationofsegmentedheartbeatsignalsisprovided.Therearegenerallyseveralstepsforclassification.Sincedirectclassifierexecutiononsegmentedsignalsisnotefficientintermsofperformance,thedistinguishingfeaturesofthesesignalsarehand-craftedandusedasinputdata.WavelettransformisoneofthecommonlyusedmethodstoobtainECGsignalfeatures[11-15].ByusingsomestatisticalmethodssuchasPrincipalComponentAnalysis（PCA）andLinearDiscriminantAnalysis（LDA）onthesecoefficient