Key points are not available for this paper at this time.
The shift of artificial intelligence for antimicrobial resistance (AI-AMR) from proof-of-concept studies to clinically embedded decision support critically hinges on establishing rigorous reproducibility, interpretability, and evaluation standards aligned with antimicrobial stewardship and patient safety. This review traces the field's evolution from rule-based gene matching through classical machine learning to deep learning models, foundation models, and generative models and proposes a pragmatic standards framework covering dataset curation, multi-site external validation, transparent model cards, and systematic error-cost analyses. Historically, progress was catalysed by curated resistome ontologies; modern practice demands FAIR-aligned curation, explicit bias and data-leakage audits, and prospective temporal and external geographic validation guided by emerging healthcare AI guidelines. To translate accuracy into safer prescribing, the review advocates cost-sensitive evaluation that quantifies false-positive and false-negative harms, integrates stewardship metrics (time-to-effective therapy, spectrum narrowing, days of therapy), and facilitates continuous post-deployment monitoring. Looking forward, federated learning, multimodal and foundation architectures, generative models for antimicrobial and peptide design, and explainable interfaces usable at the point-of-care are poised to reshape trustworthy and clinically deployable AI-AMR. The review concludes with a checklist for implementers: FAIR and diversity-by-design data curation, prospectively specified multi-site validation, standardised governance-linked model cards, and explicit stewardship-oriented error-cost trade-offs.
Sardar et al. (Fri,) studied this question.