簡介:IEEETRANSACTIONSONBIOMEDICALENGINEERING,VOL59,NO6,JUNE20121499AUTOMATICMOTIONANDNOISEARTIFACTDETECTIONINHOLTERECGDATAUSINGEMPIRICALMODEDECOMPOSITIONANDSTATISTICALAPPROACHESJINSEOKLEE,MEMBER,IEEE,DAVIDDMCMANUS,SNEHMERCHANT,ANDKIHCHON,SENIORMEMBER,IEEEABSTRACTWEPRESENTAREALTIMEMETHODFORTHEDETECTIONOFMOTIONANDNOISEMNARTIFACTS,WHICHFREQUENTLYINTERFERESWITHACCURATERHYTHMASSESSMENTWHENECGSIGNALSARECOLLECTEDFROMHOLTERMONITORSOURMNARTIFACTDETECTIONAPPROACHINVOLVESTWOSTAGESTHEFIRSTSTAGEINVOLVESTHEUSEOFTHEFIRSTORDERINTRINSICMODEFUNCTIONFIMFFROMTHEEMPIRICALMODEDECOMPOSITIONTOISOLATETHEARTIFACTS’DYNAMICSASTHEYARELARGELYCONCENTRATEDINTHEHIGHERFREQUENCIESTHESECONDSTAGEOFOURAPPROACHUSESTHREESTATISTICALMEASURESONTHEFIMFTIMESERIESTOLOOKFORCHARACTERISTICSOFRANDOMNESSANDVARIABILITY,WHICHAREHALLMARKSIGNATURESOFMNARTIFACTSTHESHANNONENTROPY,MEAN,ANDVARIANCEWETHENUSETHERECEIVER–OPERATORCHARACTERISTICSCURVEONHOLTERDATAFROM15HEALTHYSUBJECTSTODERIVETHRESHOLDVALUESASSOCIATEDWITHTHESESTATISTICALMEASURESTOSEPARATEBETWEENTHECLEANANDMNARTIFACTS’DATASEGMENTSWITHTHRESHOLDVALUESDERIVEDFROM15TRAININGDATASETS,WETESTEDOURALGORITHMSON30ADDITIONALHEALTHYSUBJECTSOURRESULTSSHOWTHATOURALGORITHMSAREABLETODETECTTHEPRESENCEOFMNARTIFACTSWITHSENSITIVITYANDSPECIFICITYOF9663AND9473,RESPECTIVELYINADDITION,WHENWEAPPLIEDOURPREVIOUSLYDEVELOPEDALGORITHMFORATRIALFIBRILLATIONAFDETECTIONONTHOSESEGMENTSTHATHAVEBEENLABELEDTOBEFREEFROMMNARTIFACTS,THESPECIFICITYINCREASEDFROM7366TO8504WITHOUTLOSSOFSENSITIVITY7448–7462ONSIXSUBJECTSDIAGNOSEDWITHAFFINALLY,THECOMPUTATIONTIMEWASLESSTHAN02SUSINGAMATLABCODE,INDICATINGTHATREALTIMEAPPLICATIONOFTHEALGORITHMSISPOSSIBLEFORHOLTERMONITORINGINDEXTERMSATRIALFIBRILLATIONAF,EMPIRICALMODEDECOMPOSITIONEMD,HOLTERRECORDING,MOTIONANDNOISEMNARTIFACTDETECTION,STATISTICALMETHODIINTRODUCTIONWEHAVERECENTLYDEVELOPEDANALGORITHMFORACCURATEANDREALTIMEDETECTIONOFATRIALFIBRILLATIONAFTHATISWELLSUITEDFORCONTINUOUSECGMONITORINGAPPLICATIONS1MANUSCRIPTRECEIVEDJANUARY21,2011REVISEDJUNE27,2011ANDOCTOBER25,2011ACCEPTEDNOVEMBER5,2011DATEOFPUBLICATIONNOVEMBER10,2011DATEOFCURRENTVERSIONMAY18,2012THISWORKWASFUNDEDINPARTBYTHEOFFICEOFNAVALRESEARCHWORKUNITUNDERGRANTN000140810244ASTERISKINDICATESCORRESPONDINGAUTHORJLEEISWITHTHEDEPARTMENTOFBIOMEDICALENGINEERING,WORCESTERPOLYTECHNICINSTITUTE,MA01609USAEMAILJINSEOKWPIEDUDDMCMANUSISWITHTHECARDIOLOGYDIVISION,DEPARTMENTSOFMEDICINEANDQUANTITATIVEHEALTHSCIENCES,UNIVERSITYOFMASSACHUSETTSMEDICALCENTER,WORCESTER,MA01605USAEMAILMCMANUSDUMMHCORGSMERCHANTISWITHTHESCOTTCARECORPORATION,CLEVELAND,OH44135USAEMAILSMERCHANTSCOTTCARECOMKHCHONISWITHTHEDEPARTMENTOFBIOMEDICALENGINEERING,WORCESTERPOLYTECHNICINSTITUTE,MA01609USAEMAILKICHONWPIEDUCOLORVERSIONSOFONEORMOREOFTHEFIGURESINTHISPAPERAREAVAILABLEONLINEATHTTP//IEEEXPLOREIEEEORGDIGITALOBJECTIDENTIFIER101109/TBME20112175729USEOFECGMONITORSEG,HOLTERMONITORSISCOMMONINTHEDIAGNOSISANDMANAGEMENTOFPATIENTSWITH,ORATRISKFOR,AF,GIVENTHEPAROXYSMAL,SHORTLIVED,ANDFREQUENTLYASYMPTOMATICNATUREOFTHISSERIOUSARRHYTHMIAMONITORINGFORAFISIMPORTANTBECAUSE,DESPITEOFTENBEINGPAROXYSMALANDASSOCIATEDWITHMINIMALORNOSYMPTOMS,AFISASSOCIATEDWITHSEVEREADVERSEHEALTHCONSEQUENCES,INCLUDINGSTROKE,HEARTFAILURE,ANDDEATH2OURTESTOFACCURACYOFTHEAFALGORITHMWASPERFORMEDONNOISEREMOVEDTESTDATABASES,WHICHALSOCONSISTEDOFHOLTERRECORDINGSCERTAINLY,MOTIONANDNOISEMNARTIFACTSARESIGNIFICANTDURINGHOLTERRECORDINGSANDCANLEADTOFALSEDETECTIONSOFAFCLINICIANSHAVECITEDMNARTIFACTSINAMBULATORYMONITORINGDEVICESASTHEMOSTCOMMONCAUSEOFFALSEALARMS,LOSSOFSIGNAL,ANDINACCURATEREADINGS3,4PREVIOUSCOMPUTATIONALEFFORTSHAVELARGELYRELIEDONMNARTIFACTREMOVAL,ANDSOMEOFTHEPOPULARMETHODSINCLUDELINEARFILTERING5,ADAPTIVEFILTERING6,7,WAVELETDENOISING8–10,ANDBAYESIANFILTERINGMETHODS11ONEMAINDISADVANTAGEOFTHEADAPTIVEFILTERINGMETHODSISTHATTHEYREQUIREAREFERENCESIGNAL,WHICHISPRESUMEDTOBECORRELATEDINSOMEWAYWITHTHEMNARTIFACTSFORMITIGATINGTHISLIMITATION,USEOFACCELEROMETERSTOOBTAINAREFERENCESIGNALHASRESULTEDINSOMESUCCESS12,13HOWEVER,THISAPPROACHHASNOTBEENAPPLIEDTOHOLTERMONITORSTHEWAVELETDENOISINGAPPROACHATTEMPTSTOSEPARATECLEANANDNOISYWAVELETCOEFFICIENTS,BUTITCANBEDIFFICULTTOUSESINCEITREQUIRESIDENTIFICATIONOFTHELOCATIONOFEACHECGMORPHOLOGYINCLUDINGTHEPANDTWAVES8–10BAYESIANFILTERINGREQUIRESESTIMATIONOFOPTIMALPARAMETERSUSINGANYVARIANTOFKALMANFILTERINGMETHODSEXTENDEDKALMANFILTER,EXTENDEDKALMANSMOOTHER,ORUNSCENTEDKALMANFILTER11THEMAINDISADVANTAGEOFTHEBAYESIANFILTERINGAPPROACHISTHEIMPROPERASSUMPTIONTHATNOISEHASANADDITIVEGAUSSIANPROBABILITYDENSITYFUNCTIONFURTHER,THEMETHODREQUIRESRPEAKLOCATIONSFOREACHCYCLEOFECGDATAWHILETHEAFOREMENTIONEDSIGNALPROCESSINGAPPROACHESHAVEBEENAPPLIED,THEYARENOTAPPROPRIATE,ANDCONSEQUENTLYMNARTIFACTSREMAINAKEYOBSTACLETOTHEACCURATEDETECTIONOFAFANDATRIALFLUTTER,WHICHISANEQUALLYPROBLEMATICARRHYTHMIAANOVELMETHODTOSEPARATECLEANECGPORTIONSFROMSEGMENTSWITHMNARTIFACTSINREALTIMEISURGENTLYNEEDEDFORMOREACCURATEDIAGNOSISANDTREATMENTOFCLINICALLYIMPORTANTATRIALARRHYTHMIASFOROURPAPER,THEAIMISTODETECTTHEPRESENCEOFMNARTIFACTSFORHOLTERAPPLICATIONS,THEREAREASUFFICIENTNUMBEROFCLEANSEGMENTSINEACHRECORDINGTHATMNCONTAMINATEDSEGMENTSCANBEDISCARDED,THEREBYINCREASINGTHESPECIFICITYOFAFIDENTIFICATIONMOREOVER,OURAFDETECTIONALGORITHMIS00189294/3100?2012IEEELEEETALAUTOMATICMOTIONANDNOISEARTIFACTDETECTIONINHOLTERECGDATA1501FIG2SQUAREDIMFBASEDONCLEANANDNOISYECGSIGNALACLEANECGSEGMENTBNOISYECGSEGMENTFIG3SIMPLIFIEDALGORITHMFORMNARTIFACTDETECTIONINANECGSEGMENTBYUSINGEMDANDTHREESTATISTICALTECHNIQUESFIG1AANDNOISYSIGNALSSEEFIG1BASSHOWNINFIG2,THEPEAKAMPLITUDESOFTHECLEANSIGNALSEEFIG2AAREANORDEROFMAGNITUDEHIGHERTHANTHOSEOFTHEMNCORRUPTEDSIGNALSEEFIG2B,INDICATINGTHATATHRESHOLDVALUECANBEDERIVEDTOSEPARATEBETWEENTHETWOTYPESOFSIGNALSWITHANORMALIZEDSQUAREDIMF,WEDETERMINETHEOPTIMUMLOWNOISELEVELTHRESHOLDLNLTVALUEANDDEFINEITASTHLNLTFOREACHTHLNLTVALUESTARTINGFROM0TO1ATANINCREMENTOF005,WEINVESTIGATETHEFOLLOWINGTHREESTATISTICALINDICESSHANNONENTROPYTOCHARACTERIZERANDOMNESS,AMEANVALUETOQUANTIFYLNLTLEVEL,ANDVARIANCETOQUANTIFYVARIABILITYIFALLVALUESOFSHANNONENTROPY,MEAN,ANDVARIANCEAREHIGHERTHANTHRESHOLDVALUESOFTHENT,THMEAN,ANDTHVAR,WEDECLARETHESEGMENTTOBEANOISECORRUPTEDSEGMENTTHEOVERALLALGORITHMISSUMMARIZEDINFIG3ONCETHLNLTANDTHETHRESHOLDSFORMAXIMUMSENSITIVITYANDSPECIFICITYAREDETERMINEDFOREACHOFTHETHREESTATISTICALVALUESTHENT,THMEAN,ANDTHVARUSINGTHERECEIVER–OPERATORCHARACTERISTICCURVEANALYSISONTHEDATA,ASDESCRIBEDINSECTIONIIB,NOFURTHERHEURISTICTUNINGFORTHETHRESHOLDVALUESISREQUIREDWEALSOINVESTIGATEDTHEOPTIMUMSEGMENTLENGTHLSEGFORMAXIMUMSENSITIVITYANDSPECIFICITYALONGWITHCOMPUTATIONALCOMPLEXITYBDATAACQUISITIONIDATACOLLECTIONANDDETERMINATIONOFOPTIMALTHRESHOLDVALUESWECOLLECTED5LEADECGHOLTERRECORDINGSSCOTTCARECORPORATIONFROM15HEALTHYSUBJECTSDATAWEREACQUIREDAT180HZWITH10BITRESOLUTIONFOR24HNONEOFTHESUBJECTSHADCLINICALLYAPPARENTCARDIOVASCULARDISEASETHE15HEALTHYSUBJECTSCOMPRISED8FEMALESAND7MALESOFAGE317±34YEARSDURINGHOLTERRECORDING,EACHSUBJECTWASASKEDTOPERFORMROUTINEDAILYACTIVITIESAMONGTHEACQUIREDDATA,WECOLLECTED14410SNOISYSEGMENTS,WHERERPEAKSWERENOTCLEARLYRECOGNIZABLEDUETOMNARTIFACTSALONGWITHTHENOISYSEGMENTS,WECOLLECTED14410SCLEANSEGMENTS,WHERERRINTERVALSWERECLEARLYDISCERNIBLENOTETHATTHEDECISIONTODEEMASEGMENTNOISECORRUPTEDORCLEANWASBASEDONTHECRITERIONOFWHETHERORNOTTHERPEAKSOFTHEECGWAVEFORMSWERERECOGNIZABLETOTHEEYEFORTHESELECTIONOFTHEOPTIMALTHRESHOLDSETCONSISTINGOFTHLNLT,THENT,THMEAN,ANDTHVAR,WESEARCHEDEVERYPOSSIBLECOMBINATIONAMONGTHE4DVECTORSWITHTHEFOLLOWINGINTERVALINCREMENTS1THLNLTVARIEDFROM0TO1ATINTERVALSOF0052THENTVARIEDFROM0TO1ATINTERVALSOF000013THMEANVARIEDFROM0TO1ATINTERVALSOF000014THRMSSDVARIEDFROM0TO001ATINTERVALSOF000001THEOPTIMALTHRESHOLDWASDETERMINEDACCORDINGTOACOMBINATIONOFTHEFOURTHRESHOLDVALUESTHATPROVIDEDTHEBESTACCURACYTHEACCURACYWASCALCULATEDASFOLLOWSACCURACYTPTNTPTNFPFN1WHERETP,TN,FPANDFNARETRUEPOSITIVES,TRUENEGATIVES,FALSEPOSITIVES,ANDFALSENEGATIVES,RESPECTIVELYWITHTHEDATALENGTH,LSEG5S,WEFOUNDTHEACCURACYOF09688,ANDTHESENSITIVITYANDSPECIFICITYVALUESOF09549AND09792,RESPECTIVELY1OPTIMALDATALENGTHANDCOMPUTATIONALTIMETODETERMINETHEOPTIMUMDATALENGTHLSEGFORMNARTIFACTSDETECTION,WEREPEATEDTHEAFOREMENTIONEDPROCEDUREWITHASEGMENTSIZEVARYINGFROM1TO10SATANINCREMENTOF1SBASEDONEACHLSEG1–10S,WEOBTAINEDTHEOPTIMALPARAMETERSEG,10SETSOFTHRESHOLDSETSANDPLOTTEDTHEACCURACYACCORDINGTOLSEG,ASSHOWNINFIG4ATHEACCURACYINCREASEDWHENLSEGINCREASED,BUTTHERATEOFINCREASEDECLINEDWHENLSEGWASEQUALTOORGREATERTHAN5SINADDITION,ASSHOWNINFIG4B,THECOMPUTATIONTIMEFORACLEANSEGMENTLINEARLYINCREASEDWITHTHELENGTHOFDATASEGMENTSHOWEVER,THECOMPUTATIONTIMEFORNOISYSEGMENTSDRAMATICALLYINCREASEDESPECIALLYWHENTHESEGMENTLENGTHEXCEEDED6S,ASSHOWNINFIG4CTAKINGINTOACCOUNTTHECOMPUTATIONALCOMPLEXITY,WECHOSETHEOPTIMUMLSEG5SNOTETHATTHECOMPUTATIONALTIMEWASOBTAINEDBYMATLAB2010AON266GHZINTELCORE2PROCESSORTABLEISUMMARIZESTHEFINALOPTIMALTHRESHOLDPARAMETERSAND
下載積分: 10 賞幣
上傳時間:2024-03-13
頁數(shù): 8
大?。?0.61(MB)
子文件數(shù):