Why Artificial Intelligence Models for Antimicrobial Resistance Still Fail at the Bedside: A Review of Validation, Explainability, and Equity Gaps

Abstract

Artificial intelligence (AI) and machine learning (ML) models now predict antimicrobial resistance (AMR) phenotypes with striking accuracy in retrospective, single-centre studies — and yet, curiously, almost none of them have made it as far as routine clinical use. This review set out to understand that gap, and, more usefully, what might close it. Following PRISMA 2020 guidance, we systematically synthesised peer-reviewed and preprint literature (2017–2026) on AI/ML applications for AMR prediction, mapping studies across four domains: comparative model performance, diagnostic input modality, data heterogeneity and algorithmic bias, and technological readiness in low- and middle-income countries (LMICs). Risk of bias was appraised using a PROBAST-informed framework. Tree-based ensembles and deep learning architectures achieved internal AUROC values of 0.89–0.99 across genomic and mass-spectrometric datasets; externally validated cohorts, however, showed a consistent 0.10–0.25 AUROC decay. Over 80% of publicly available pathogen genomic data originate from North America and Europe, while African isolates contribute under 2%. Calibration reporting and biologically validated explainability were almost entirely absent from the reviewed literature, and One Health data integration remained confined to fewer than 1% of studies. Taken together, these findings suggest that the translational gap in AI-based AMR prediction is not, at its core, a modelling problem — it is a problem of representativeness, validation rigour, and earned clinical trust. Closing it will require external and prospective validation as a reporting norm, calibrated and biologically grounded explainability, federated infrastructure inclusive of LMIC data, and tiered, offline-capable deployment pathways suited to resource-limited settings.