Explainable Deep Learning Approach for Multi label Classification of Antimicrobial Resistance with Missing Labels

Abstract

PredictingAntimicrobialResistance(AMR)fromgenomicsequencedatahasbecomea significantcomponentofovercomingtheAMRchallenge,especiallygivenitspotentialforfacilitatingmore rapiddiagnosticsandpersonalisedantibiotictreatments.Withtherecentadvancesinsequencingtechnologies andcomputingpower,deeplearningmodelsforgenomicsequencedatahavebeenwidelyadoptedtopredict AMRmorereliablyanderror-free.TherearemanydifferenttypesofAMR;therefore,anypracticalAMR predictionsystemmustbeabletoidentifymultipleAMRspresentinagenomicsequence.Unfortunately, mostgenomicsequencedatasetsdonothaveallthelabelsmarked,therebymakingadeeplearningmodelling approachchallengingowingtoitsrelianceonlabelsforreliabilityandaccuracy.Thispaperaddresses thisissuebypresentinganeffectivedeeplearningsolution,Mask-Loss1Dconvolutionneuralnetwork (ML-ConvNet),forAMRpredictionondatasetswithmanymissinglabels.Thecorecomponentof ML-ConvNetutilisesamaskedlossfunctionthatovercomestheeffectofmissinglabelsinpredicting AMR.TheproposedML-ConvNetisdemonstratedtooutperformstate-of-the-artmethodsintheliteratureby 10.5%,accordingtotheF1score.Theproposedmodel’sperformanceisevaluatedusingdifferentdegrees ofthemissinglabelandisfoundtooutperformtheconventionalapproachby76%intheF1scorewhen 86.68%oflabelsaremissing.Furthermore,theML-ConvNetwasestablishedwithanexplainableartificial intelligence(XAI)pipeline,therebymakingitideallysuitedforhospitalandhealthcaresettings,wheremodel interpretabilityisanessentialrequirement.