Comparison of OneChoice AI-Based Clinical Decision Support Recommendations with Infectious Disease Specialists and Non-Specialists for Empirical Urinary Tract Infection Therapy in Lima, Peru

Background: Antimicrobial resistance complicates the selection of appropriate regimens for urinary tract infections (UTIs), even when susceptibility data are available, particularly where infectious disease (ID) expertise is scarce. Machine learning clinical decision support systems (CDSS) may support prescribing, but evidence from Latin America is limited. The goal of this study was to evaluate the concordance between antimicrobial regimens selected by physicians and those recommended by a machine-learning-with-human-in-the-loop (ML-HITL) CDSS (OneChoice®), and to assess CDSS appropriateness against an independent, blinded expert reference standard. Methods: In this cross-sectional, survey-based concordance study conducted in Lima, Peru, 194 verified physicians contributed 224 eligible evaluations across 42 real UTI case codes with complete culture and antimicrobial susceptibility data. Of the 224 evaluations, 70 were contributed by infectious disease specialists and 154 by non-ID physicians. Participants selected OneChoice® and alternative antimicrobial regimens. Responses were compared with CDSS recommendations under three concordance definitions. Discordances were adjudicated by an external panel blinded to the source of the recommendation. Non-independence was addressed using cluster-robust methods. Results: First-choice, alternative, and general concordance were 50.9%, 40.6%, and 62.5%, respectively. ID specialists showed higher concordance than non-ID physicians (65.7% vs. 44.2%; 51.4% vs. 35.7%; 72.9% vs. 57.8%; all p ≤ 0.034). ID specialty was independently associated with concordance (adjusted OR 2.19–2.68; p ≤ 0.016). Among discordant evaluations, the external panel judged the CDSS recommendation to be preferable in 90.5–93.2% of cases. Physician–CDSS concordance was moderate and higher among ID specialists. The external adjudication findings indicate that the CDSS recommendations were frequently aligned with expert assessment when physician and CDSS recommendations differed; however, the study did not evaluate comparative clinical effectiveness or patient outcomes. Conclusions: Physician–CDSS concordance was moderate and higher among ID specialists, yet discordances overwhelmingly favored the CDSS on independent adjudication. These findings suggest the CDSS aligns with expert reasoning and may support antimicrobial selection in high-resistance settings.