Pattern recognition using collaborative neural networks

Abstract

Artificial neuralnetworkshavebeensuccessfullyappliedtoawiderangeofproblems.Inpat- tern recognition,theyhavebeenusedinseveraltasks,suchasfeatureextraction,dimension reduction, andclassification.Inthiswork,weproposetwoECGheartbeatclassificationmodels based oncollaboratingdifferenttypesofneuralnetworks.Themainaimistocombinetheir complementaryproperties. The firstmodelusesastackedsparseautoencoder(SSAE)asfeatureextractorandasystem of multipleMulti-layeredperceptrons(MLP)asaclassifier.Inthismodel,theentireproblem is dividedintosimplerparts,whichareresolvedusingdifferentMLPs.Thesecondmodelalso uses aSSAEtoextractfeaturesinadditiontotwootherdynamicfeatures.Inthismodel,the classification isperformedbyahybridneuralmodelbasedoncombiningrandomandRBFneural networks. The proposedmodelsareevaluatedontheMIT-BIHarrhythmiadataset.Thetestsarebasedon the inter-patientparadigm,inwhichthetrainingandtestdataaretakenfromdifferentpatients. The obtainedresultsarecomparedwithsomeofthestate-of-the-artmethods.