AI models for antimicrobial resistance are moving from promising classifiers to stewardship tools
Artificial intelligence for antimicrobial resistance is no longer limited to isolated proof-of-concept studies or retrospective benchmark reports. The field now faces a harder test: whether AI-AMR can support prescribing decisions without creating new risks through opaque predictions, hidden data leakage, or performance that collapses outside the dataset where a model was born.
The review tracks this progression from rule-based gene matching to classical machine learning, then onward to deep learning, foundation models, and generative systems. Each step has expanded technical capability, but the central problem has stayed the same. Resistance prediction is only clinically useful if it is reproducible, interpretable, and tied to the consequences of treatment delay, unnecessary broad-spectrum therapy, and missed active agents.
A model that looks accurate on paper can still be unsafe at the bedside.
Why reproducibility now sits at the center of AI-AMR
Earlier resistance tools depended heavily on curated resistome ontologies and gene-to-phenotype rules. Those resources helped establish the field, but they also exposed a structural weakness: many models were trained in narrowly defined settings with limited diversity in organisms, institutions, and sequencing or phenotypic methods. Once such systems are moved into different hospitals or different time periods, hidden biases can surface quickly.
This review argues for a more rigorous framework built around FAIR-aligned curation, meaning data must be findable, accessible, interoperable, and reusable. The practical effect is straightforward. If training data are incomplete, poorly annotated, or assembled from overlapping sources, even a technically elegant model may learn the wrong signals. Leakage between training and validation sets can inflate performance in ways that are invisible until deployment fails. In antimicrobial resistance, where a false prediction can expose a patient to ineffective therapy or unnecessary escalation, those errors are not abstract statistical defects; they are clinical events.
External validation across sites matters because resistance patterns are local, temporal, and organism-specific. A model trained in one region may not reflect another hospital’s antibiogram, infection control pressure, or sequencing workflow. Temporal validation is equally important because resistance evolves under antimicrobial selection pressure. A system that performs well on older isolates may miss emerging phenotypes when resistance mechanisms shift.
Interpretability has to work at the point of care, not just in a methods section
Explainability is often discussed as a technical feature, but the clinical demand is more concrete. A prescriber needs to know whether a resistant call reflects a known mechanism, an uncertain borderline phenotype, or a model pattern driven by confounding site-specific data. Without that context, the output may be ignored or misapplied.
The review supports transparent model cards and governance-linked reporting so that users can see what a system was trained on, where it has been tested, how it behaves across subgroups, and what kinds of errors it tends to make. In antimicrobial stewardship, this is not a cosmetic addition. It is the difference between a tool that can be audited and one that merely produces probabilities.
Error costs matter more than raw accuracy in resistance prediction
Accuracy alone is an incomplete endpoint for AMR prediction. False positives and false negatives carry different harms, and those harms depend on the clinical context. A false positive resistance call may push clinicians toward broader therapy than needed, increasing toxicity, cost, and selection pressure. A false negative may delay effective treatment and prolong infection before the correct drug is started.
The review calls for cost-sensitive evaluation that quantifies those trade-offs directly. This is a stronger standard than reporting overall correctness because it forces the model to be judged by the consequences of its mistakes. Stewardship-oriented outcomes should also be built into evaluation, including time-to-effective therapy, spectrum narrowing, and days of therapy. Those metrics connect algorithm performance to the workflow of infectious diseases care instead of leaving the result trapped inside a ROC curve.
In practical terms, a resistance model that saves minutes but increases unnecessary carbapenem use has failed its clinical mission.
Post-deployment monitoring is not optional for a changing microbiology field
Antimicrobial resistance evolves after deployment, which means the model’s environment changes while the model itself remains fixed unless actively maintained. Continuous monitoring therefore becomes a safety requirement rather than an afterthought. Performance drift can arise from changes in organism prevalence, local resistance mechanisms, patient mix, specimen type, laboratory practice, or antibiotic prescribing behavior.
The review supports ongoing surveillance after deployment, with systematic review of prediction errors and governance processes that can trigger retraining or recalibration. This mirrors how stewardship programs already track resistance trends and antibiotic utilization. The same logic should apply to AI systems that are meant to influence those decisions.
Without post-deployment monitoring, a resistance model can become stale while still appearing authoritative.
Federated learning, multimodal systems, and generative design may widen the field
Several newer architectures are positioned as next steps. Federated learning could allow multi-site model development without centralizing sensitive patient data, a useful approach for hospitals that cannot easily share raw microbiology records. Multimodal models may combine genomics, phenotypes, clinical context, and local epidemiology, which is attractive because resistance is rarely explained by a single data stream. Foundation models may offer broader representational capacity, while generative models could support antimicrobial and peptide design rather than prediction alone.
Those advances are technically exciting, but they also magnify the same governance demands. Larger models can be harder to audit, and generative systems can introduce new questions about safety, novelty, and unintended biological effects. Explainable interfaces usable at the bedside will be essential if these tools are to influence prescribing rather than remain research infrastructure.
What implementers must build before clinical deployment
The review ends with a checklist that is less about innovation theater and more about operational discipline. Data curation should reflect diversity by design, not as a retrospective correction. Validation should be multi-site and prospectively specified. Model cards should be standardized and linked to governance. Error-cost trade-offs should be explicit, not implied.
That framework is demanding because AI-AMR sits at the intersection of microbiology, informatics, and stewardship. Success will not come from higher accuracy alone. It will come from models that can be trusted in real time, across changing hospital ecologies, with transparent failure modes and measurable effects on antibiotic use.
The unresolved challenge is not whether these systems can classify resistance in controlled settings. It is whether they can keep doing so after they meet the complexity of live clinical practice, where organisms evolve, workflows vary, and the cost of a wrong answer is immediate.