Trans AMR: An Interpretable Transformer Model for Accurate Prediction of Antimicrobial Resistance using Antibiotic Administration Data

Abstract

AntimicrobialResistance(AMR)isagrowingpublicandveterinaryhealthconcern,andthe abilitytoaccuratelypredictAMRfromantibioticsadministrationdataiscrucialforeffectivelytreatingand managinginfections.Whilegenomics-basedapproachescanprovidebetterresults,sequencing,assembling, andapplyingMachineLearning(ML)methodscantakeseveralhours.Therefore,alternativeapproachesare required.ThisstudyfocusedonusingMLforantimicrobialstewardshipbyutilisingdataextractedfrom hospitalelectronichealthrecords,whichcanbedoneinreal-time,anddevelopinganinterpretable1DTransformermodelforpredictingAMR.Amulti-baselineIntegratedGradientpipelinewasalsoincorporated tointerpretthemodel,andquantitativevalidationmetricswereintroducedtovalidatethemodel.The performanceoftheproposed1D-Transformermodelwasevaluatedusingadatasetofurinarytractinfection (UTI)patientswithfourantibiotics.Theproposed1D-Transformermodelachieved10%higherareaunder curve(AUC)inpredictingAMRandoutperformedtraditionalMLmodels.TheExplainableArtificial Intelligence(XAI)pipelinealsoprovidedinterpretableresults,identifyingthesignaturescontributingto thepredictions.Thiscouldbeusedasadecisionsupporttoolforpersonalisedtreatment,introducing AMR-awarefoodandmanagementofAMR,anditcouldalsobeusedtoidentifysignaturesfortargeted interventions.