AI-based clinical decision support systems for antibiotic therapy in common adult infections: a systematic review from development to clinical implementation

Abstract

Antimicrobial resistance is among the leading causes of global mortality, and artificial intelligence (AI)-based clinical decision support systems (CDSS) have been proposed to guide antibiotic prescribing. Several AI models for antibiotic decision-making have been developed and validated, but their actual deployment in clinical practice and effect on patient outcomes have not been systematically characterized. To identify, classify, and appraise studies evaluating AI-based CDSS for antibiotic therapy in hospitalized adults with common infections, stratified by level of clinical implementation. Search-strategy development began on June 7, 2026; definitive searches of PubMed/MEDLINE, Scopus, Web of Science Core Collection, and Embase were executed on June 28, 2026, after submission of the PROSPERO record. Four reviewers independently screened records and assessed reports. Studies were classified as deployed in clinical practice (Level 1), evaluated against physicians' decisions without deployment (Level 2), or tested through retrospective simulation (Level 3). Risk of bias was assessed with RoB 2, ROBINS-I, or PROBAST, as appropriate. Owing to clinical and methodological heterogeneity, findings were synthesized narratively, with an outcome-specific GRADE assessment. Of 1,157 records identified, 10 studies met the inclusion criteria: three Level 1 (30%), four Level 2 (40%), and three Level 3 (30%). The three deployed studies evaluated pneumonia, predominantly pulmonary Stenotrophomonas maltophilia infection, and Staphylococcus aureus bacteremia and reported favorable signals for mortality, hospital processes, prescribing, or medication safety. Kanjilal et al., a published conference abstract, was included as a Level 3 retrospective evaluation of 29,508 adults with community-onset sepsis. Among the seven Level 2/3 studies assessed with PROBAST, three were classified as low, three as moderate under the prespecified descriptive adaptation, and one as high overall risk of bias. Certainty was low for mortality/survival and very low for the remaining clinically relevant outcomes. Three of 10 included studies reached clinical deployment, whereas seven remained non-deployed. Deployed systems showed favorable outcome and process signals, but the evidence was heterogeneous and of low to very-low certainty, depending on the outcome. Prospective multicenter evaluations of deployed systems are required before computational performance can be interpreted as clinical effectiveness. https://www.crd.york.ac.uk/PROSPERO/view/CRD420261435593 , identifier CRD420261435593.