AI-Driven Multidrug-Resistance Prediction Pilot — Jamhuriya University Tests Random-Forest Triage for Mogadishu Hospitals

Somalia has among the weakest antimicrobial-resistance (AMR) surveillance infrastructure in the world, with laboratory capacity and consistent reporting still being built up nationally. Researchers at Jamhuriya University of Sciences and Technology's Center for Antimicrobial Resistance Research set out to test whether a lightweight machine-learning model, trained only on data routinely captured at the point of care, could help flag likely multidrug-resistant (MDR) infections in settings without full genomic or advanced laboratory capacity.
From a pool of 1,012 bacterial isolates collected at sentinel healthcare facilities in Mogadishu between September 2024 and January 2026, the team built a random-forest classifier on a final analytic cohort of 262 isolates, using only patient age, sex, organism type, hospital ward and specimen source as predictors of resistance to five sentinel antibiotics. On a held-out test set of 51 isolates, the model reached 66.7% overall accuracy (95% CI 52.1-79.2%) and an AUC of 0.607. Sensitivity was high at 91.4%, meaning it caught most true MDR cases, but specificity was just 12.5% — it wrongly flagged the large majority of non-resistant isolates as resistant too.
The authors are explicit that this is a "preliminary proof-of-concept" rather than a deployment-ready tool: the analytic cohort was a quarter of the original pool, with MDR prevalence enriched by selection bias (68.3% versus 32.6% in the full pool); the study was single-site; and no external validation, cross-validation or hyperparameter tuning was performed. Evidoria includes it as an honest, well-documented example of the gap between a promising pilot and a trustworthy public-health tool — useful precisely because it shows what still needs fixing before an AI triage aid can be trusted in a low-resource clinical setting.

Read the full analysis: https://www.dovepress.com/an-ai-driven-antimicrobial-resistance-surveillance-framework-for-low-r-peer-reviewed-fulltext-article-IDR