Artificial Intelligence in Antimicrobial Stewardship: Transforming Antibiotic Prescribing in the Digital Era

Abstract

Background: Antimicrobial resistance (AMR) is a major global health threat, driven in part by inappropriate antibiotic prescribing, and clinical workload, workforce shortages, and fragmented data systems increasingly strain conventional antimicrobial stewardship (AMS). Artificial intelligence (AI) is proposed as a means of augmenting AMS by integrating heterogeneous clinical, microbiological, and pharmacy data. Objective: To critically synthesize the evidence for AI-enabled AMS across the antibiotic-prescribing pathway, appraise the maturity of that evidence, and identify the barriers separating algorithmic performance from clinical benefit. Methods: Narrative review. PubMed/MEDLINE, Google Scholar, and preprint/repository sources were searched for English-language records published between 2020 and 2026 using combinations of “artificial intelligence,” “machine learning,” “antimicrobial stewardship,” “antibiotic prescribing,” “clinical decision support,” and “antimicrobial resistance,” supplemented by hand-searching reference lists of identified reviews. Sources were prioritized by relevance and study type, with a preference for systematic reviews, meta-analyses, and primary clinical evaluations over case reports or non-peer-reviewed material. Results: AI applications span infection prediction, empiric antibiotic selection, resistance prediction, dose and duration optimization, and de-escalation. Predictive performance is often encouraging (pooled AUC ≈72% across 80 studies in one meta-analysis), but the evidence is dominated by small, single-center, retrospective studies with limited external validation; only a handful of prospective or randomized evaluations exist, and none demonstrate consistent, generalizable improvement in patient-level or stewardship outcomes. Explainability, workflow integration, alert fatigue, governance, and infrastructure and data representativeness, particularly in low- and middle-income countries (LMICs), remain unresolved and are rarely addressed jointly with model development. Conclusions: Current evidence supports AI as a promising but clinically unproven adjunct to AMS. The literature’s emphasis on algorithmic accuracy over demonstrated clinical effectiveness means AI should not yet be treated as validated stewardship infrastructure; prospective, externally validated, multicenter evaluation, not further retrospective modelbuilding, is the priority for the field.