Feature-determination-for-heart-sounds-based-on-di

DigitalSignalProcessing19(2009)521–531ümrayDokur∗,TamerÖlmezDepartmentofElectronicsandCommunicationEngineering,IstanbulTechnicalUniversity,34469Istanbul,TurkeyAvailableonline26December2007AbstractHeartauscultation(theinterpretationofheartsoundsbyaphysician)isafundamentalcomponentofcardiacdiagnosis.Itis,however,adifficultskilltoacquire.Indecisionmaking,itisimportanttoanalyzeheartsoundsbyanalgorithmtogivesupporttomedicaldoctors.Inthisstudy,twofeatureextractionmethodsarecomparativelyexaminedtorepresentdifferentheartsound(HS)categories.First,arectangularwindowisformedsothatoneperiodofHSiscontainedinthiswindow.Then,thewindowedtimesamplesarenormalized.DiscretewavelettransformisappliedtothiswindowedoneperiodofHS.Basedonthewaveletdetailcoefficientsatseveralbands,thetimelocationsofS1–S2soundsaredeterminedbyanadaptivepeakdetector.Inthefirstfeatureextractionmethod,sub-bandsbelongingtothedetailcoefficientsarepartitionedintotensegments.Powersofthedetailcoefficientsineachsegmentarecomputed.Inthesecondfeatureextractionmethod,thepowerofthesignalinawindowwhichconsistsof64samplesiscomputedwithoutfilteringtheHSs.Inthestudy,performancesofthesetwofeatureextractionmethodsarecomparativelyexaminedbythedivergenceanalysis.Theanalysisquantitativelymeasuresthedistributionofvectorsinthefeaturespace.©2007ElsevierInc.Allrightsreserved.Keywords:Heartsoundanalysis;Featureextractionforheartsounds;Wavelettransform;SegmentationofS1–S2sounds;Divergenceanalysis1.IntroductionAuscultationisatechniqueinwhichastethoscopeisusedtolistentothesoundsofabody.Thestructuraldefectsoftheheartareoftenreflectedinthesoundstheheartproduces.Physiciansusethestethoscopeasadevicetolistentothepatient’sheartandmakeadiagnosisaccordingly.Theyareparticularlyinterestedinabnormalsounds,whichmaysuggestthepresenceofacardiacpathologyandalsoprovidediagnosticinformation.Forinstance,averyimportanttypeofabnormalsoundisthemurmur,whichisasoundcausedbytheturbulentflowofbloodinthecardiovascularsystem.Thetimingandpitchofamurmurareofsignificantimportanceinthediagnosisofaheartcondition,forexam-ple,murmursduringdiastolearesignsofmalfunctioningofheartvalvesbutmurmursduringsystolemaycorrespondtoeitherapathologicalorhealthyheart,dependingontheacousticcharacteristicsofthemurmurs.Intheliterature,itisobservedthattime–frequency/scalemethodshavebeenappliedtocharacterizeheartsounds[1,2].Inpreviouspublications,theauthorshavediscussedthecharacterizationofheartmurmursusingtime–frequencymethodsoveranumberofcardiaccycles[3,4].Theacousticsignalsfromtheheartcontaininformationwhichcannotbeanalyzedbythehumanear[5].Thesensitivityoftheearinregardtofrequencyfollowsalogarithmicscale.Theearhearschangesinfrequencybetterthan*Correspondingauthor.Fax:[phone].E-mailaddresses:[e-mail](Z.Dokur),[e-mail](T.Ölmez).1051-2004/$–seefrontmatter©2007ElsevierInc.Allrightsreserved.doi:10.1016/j.dsp.2007.11.003522Z.Dokur,T.Ölmez/DigitalSignalProcessing19(2009)521–531Fig.1.Anormalheartbeat(top),withS1andS2(bottom)[26].changesinintensity.Soundswithhigherfrequenciesareperceivedasbeinglouderthanthosewithlowerfrequenciesofsameintensity.Changesinfrequencymaybeinterpretedaschangesinintensity.Inthepresenceofhigh-frequencysounds,theearmaybeunabletodetectlow-frequencyoneswhichfollowimmediately[6].Recentadvancesindatarecordingtechnologyanddigitalsignalprocessinghavemadeitpossibletorecordandanalyzethesoundsignalsfromtheheart.However,forcomputeranalysisoftheacousticsignalsfromtheheart,itisessentialthatdifferentcomponentsofheartcyclecanbetimedandseparated[7].Recentadvancesininformationtechnologysystems,indigitalelectronicstethoscopes,inacousticsignalprocessingandinpatternrecognitionmethodshaveinspiredthedesignofsystemsbasedonelectronicstethoscopesandcomput-ers[8,9].Inthelastdecade,manyresearchactivitieswereconductedconcerningautomatedandsemi-automatedheartsounddiagnosis,regardingitasachallengingandpromisingsubject.Manyresearchershaveconductedresearchonthesegmentationoftheheartsoundintoheartcycles[10–12],thediscriminationofthefirstfromthesecondheartsound[13],theanalysisofthefirstandthesecondheartsoundsandtheheartmurmurs[14–17],andalsoonfeatureextractionandclassificationofheartsoundsandmurmurs[18–23].Intheliterature,itisobservedthatwavelettransformshavebeenfrequentlyusedtoextractfeaturesfromheartsounds[20–23].Andrisevicetal.determinedthefeaturesofheartsoundsbyusingwavelettransformandprinciplecomponentanalysis[20].Theheartsoundswereclassifiedintotwocategoriesbyaneuralnetworkwithaspecificityof70.5%andasensitivityof64.7%.Guptaetal.determinedthefeaturesofheartsoundsbyusingwavelettransform[21].HeartsoundswereclassifiedintothreecategoriesbyGrowandLearnnetworkwithatotalperformanceof96%.Uguzetal.determinedthefeaturesofheartsoundsbyusingwavelettransformandshort-timeFouriertransform[22].TheheartsoundswereclassifiedintotwocategoriesbyahiddenMarkovmodelwithaspecificityof92%andasensitivityof97%.Comaketal.analyzedtheDopplersignalsofheartvalvesbyusingwavelettransfor