Artificial Intelligence in the Fight Against Antimicrobial Resistance: Applications in Diagnosis, Susceptibility Prediction, Precision Dosing, and Antibiotic Discovery

Background: Antimicrobial resistance (AMR) is a major global health threat, while conventional diagnostic, stewardship, dosing, and drug-development strategies remain limited by diagnostic delays, population-level decision-making, and a shrinking antibiotic pipeline. Artificial intelligence (AI) may help address these gaps by integrating complex clinical, microbiological, genomic, pharmacological, and chemical data. Methods: We conducted a narrative review of English-language publications issued between 1 January 2010 and 27 August 2026. PubMed/MEDLINE, Scopus, and Web of Science were searched using a prespecified combination of terms related to AI, machine learning, AMR, diagnosis, susceptibility prediction, antimicrobial stewardship, precision dosing, pharmacokinetics, pharmacodynamics, and antibiotic discovery. A separate structured search of ClinicalTrials.gov, the World Health Organization International Clinical Trials Registry Platform, and ISRCTN was performed on 27 August 2026 to identify prospective interventional studies. Results: AI-based systems demonstrated promising performance in differentiating bacterial from viral infections, predicting bacteremia and sepsis, enhancing matrix-assisted laser desorption/ionization time-of-flight interpretation, estimating antimicrobial susceptibility, supporting individualized empirical therapy, optimizing antibiotic dosing, and accelerating compound discovery. However, external and temporal validation frequently produced substantially poorer performance than internal validation. In representative resistance-prediction studies, AUROC declined from 0.94 internally to 0.55 during temporally independent external validation and by 0.10–0.25 over 18 months. Moreover, the structured registry search identified five prospective interventional studies of AI-based applications and one additional randomized study of conventional model-informed precision dosing. Conclusions: Current evidence demonstrates that AI can generate technically promising predictions, but it does not yet establish consistent clinical or economic benefit. Discrimination alone is insufficient to demonstrate clinical usefulness. Prospective multicenter trials, clinically meaningful outcome measures, external validation, continuous monitoring for model drift, and formal implementation and cost-effectiveness evaluations are required before widespread adoption.