硕士论文-基于贝叶斯网络的脉象诊断研究

上海交通大学硕士学位论文基于贝叶斯网络的脉象诊断研究姓名：肖卓申请学位级别：硕士专业：模式识别与智能系统指导教师：周越20070101III,FuzzyTAN,1peaks2IV3456TANTANTANV1peaks23VICOMPUTERIZEDPULSEANALYSISANDDIAGNOSISBASEDONBAYESIANNETWORKSABSTRACTPulseDiagnosisiswidelyusedasanimportantdiagnosticmethodinTraditionalChineseMedicineTCM.However,therearestillseveralinevitablelimitationsthatgreatlyobstructtheapplicationoftraditionalpulsediagnosticinclinicalmedicine.Firstly,pulsestatesaredescribedasakindofspecialfeelingonthefingertipsoftheexaminer,whichcan’tbewellunderstoodinwesternmedicine.Secondly,pulsediagnosisrequireslotsofexperiencesandtheinterpretationissubjective,dependingonthepractitioner.Therefore,thediagnosticresultsmaybeunreliableandinconsistent.Inthispaper,weproposeanovelcomputerizedpulseinspectionmethodaimingtoaddresstheseproblems.Firstofall,severalkeyfeaturesareextractedfromthepulsesignals.Then,Bayesiannetworksareemployedtomodeltherelationshipsbetweenthesefeaturesaswellasclassifydifferentpulsestates.Furthermore,wesetupanewTree-AugmentedNaïveBayesianClassifiermodelbasedonfuzzypartition,whichisprovedeffectiveindealingwithbothdiscreteandcontinuousvariables.VIITheresearchandinnovationinthispapermainlyincludedarelistasfollows:1)Aftersomepreprocessingproceduressuchasfiltering,smoothing,rectificationandnormalizing,weextractsevenkeyfeaturesfromthepulsesignal.Otherwise,anewfeaturenamedpeaksisinducedtodescribethegeneralshapeofapulsewave.2)ThecontinuouspulsefeaturesarediscretizedpriortoconstructionofdiscreteBayesianNetworks.Thensixkindsofpulsestates,includingbothsingle-typepulsestatesandmulti-typepulsestates,areclassifiedviaBayesianNetworkmodelslearnedbyvariousstructurelearningalgorithms.3)Consideringsomecomplicatedcasessuchasmulti-typepulsestates,weseparatethefeaturesintotwopartsaccordingtodifferentpeaksandconstructsubsidiaryBayesiannetworksrespectively.Wecangetmorereasonableandflexiblediagnosticconclusionaboutthepulsestateswithcombiningtheresultsofsubsidiarynetworks.4)ThispaperperformedaconditionalGaussianBayesiannetworkwhichassumescontinuousdataaresampledfromamultivariatenormaldistribution.Inthisnetworkmodel,learningcanbedonedirectlywithcontinuouspulsedatawithoutcommittingtoaspecificdiscretizationofthepulseattributes.ByusingGaussianstomodelthedata,wecaninduceamodelwhichapproximatethedatawellandavoidtheobviouslossofVIIIinformationcausedbydiscretization.5)WeproposeanonparametricapproachbasedonLagrange’smeanvaluetheorem.Thismethodsuitsallkindsofcontinuousvariables,includingthedatawhichisnotdistributedaccordingtoGaussiandistribution.6)Inthispaper,weextendTAN(TreeAugmentedNaïveBayesianNetwork)todealwithcontinuousattributesdirectly.Theresultisaclassifierthatcanrepresentandcombinebothdiscreteandcontinuousattributes.Inaddition,weproposeanewmethodthattakesadvantageofthemodelinglanguageofdiscreteandcontinuousformsimultaneously,andusebothversionsintheclassificationtask.Weusefuzzypartitiontogetthediscreteversion,whichcanbringsmoothnessandrobustnessinthenetworkperformance.TheempiricalresultsonUCIdatasetshowthatthislattermethodusuallyachievesclassificationperformancethatisasgoodasorbetterthaneitherthepurelydiscreteorthepurelycontinuousTANmodels.OurexperimentsonPulsedataalsoreflecttheadvantagesofthenewclassifierviaitsstableandidealclassificationratesonbothsingle-typepulsestatesandmulti-typepulsestates.KEYWORDS:TraditionalChineseMedicine,Bayesiannetwork,computerizedpulsediagnosis,continuousvariables,fuzzypartition,TANI2007126II2007126200712611.1[1][2][1][3,4]()......,21.21.2.1[5],[6][7,8]12[9-11]1.2.211[12-19]()2[20-26]Fourier33[27-29]2[30]()8012(ANN),ANN[4,31]41.31.3.1[32]Pearl1986[66-72]201NPhard2Heckerman[33]3SPI54OfficeWindowsQMR-DT6004000[34]123451.3.261.47FuzzyTANTANFuzzyTAN82.1[35-38]2.1.1[36]2-105010015020000.20.40.60.81h1h2h3h5w1w2w3th1th22-1Figure2-1Typicalpulsewavewithkeyfeatures2-1[38]9[36-39]2-1hl:h2:h3:h5:th1:,th2:,w1:1/310w2:w3:2.1.2,[1,2]12345SVMFCM[8],112-2010020030040000.51(a)010020030000.51(b)05010015020000.51(c)2-2abcFigure2-2Varioustypesofpulse:(a)typicalwirypulse,(b)typicalsmoothpulse,(c)typicalnormalpulse2-31232-42-3(1-3)(3-1)1/3w1h212010020000.51(1-2)020040000.51(1-3)(1-1)020040000.51(1-4)020040000.51020040000.51(2-1)020040000.51(2-2)020040000.51(2-3)020040000.51(2-4)010020000.51(3-1)020040000.51(3-2)020040000.51(3-3)020040000.51(3-4)2-3Figure2-3Thevarietyandflexibilityofpulsestate:differentpulsetermsinthesametype05010015020025000.20.40.60.8105010015020025000.20.40.60.8105010015020025000.20.40.60.81(a)(b)(c)2-4abcFigure2-4Multi-typepulsestates:(a)normal-smoothpulse,(b)normal-warypulse,(c)wary-smoothpulse2.22.2.1FFT131232.2.22-505010015000.51YX(xi,yi)y'2-5Figure2-5TherectificationofpulsesignalXY/kYX=k'y(2-1)'iiiyyxk=−∗(2-1)2.2.3142.1.1w1w2h2h3th2tw1th1w3h5peakspeakspeaks=1peaks=2peaks=3[40-46],,,,,,,,12342-6,151WW0.0349512HH0.0502753THTH0.770835101520253035404550050010001500GenerationFitnessvalue12300.20.40.60.8Numberofvariables(3)CurrentbestindividualCurrentBestIndividualBestfitnessMeanfitness2-6Figure2-6Theexperimentalresultsofgeneticalgorithms12iPiP'()0ifP=2121