An AI-Driven Antimicrobial Resistance Surveillance Framework for Low-Resource Settings: A Pilot Study from Mogadishu, Somalia

Abstract

Background Antimicrobial resistance (AMR) constitutes a profound global health crisis, with profoundly disproportionate impacts in low-resource settings (LRS) and conflict-affected regions like Somalia. Traditional laboratory-based surveillance is largely unfeasible in these environments due to resource constraints. This study proposes and preliminarily evaluates an Artificial Intelligence (AI)-driven surveillance framework to predict multidrug resistance (MDR) using minimal, routinely available clinical metadata. Methods A retrospective cross-sectional analysis was conducted using 262 curated clinical bacterial isolates from Mogadishu, Somalia. A Random Forest machine learning algorithm was trained using five basic variables (patient age, sex, isolated organism, ward, and specimen source) to predict MDR against five sentinel antibiotics. Phenotypic co-resistance was further evaluated using an undirected weighted network analysis. Results The predictive model achieved an overall accuracy of 66.7% (95% CI: 52.1–79.2%) and an Area Under the Curve (AUC) of 0.607. Prioritizing its utility as an early-warning triage tool, the model demonstrated an exceptionally high sensitivity (recall) of 91.4%, effectively flagging the vast majority of MDR cases, with a specificity of 12.5%, a precision of 69.6%, an F1-score of 0.79, and a Matthews Correlation Coefficient (MCC) of 0.06. The low MCC indicates that, despite high sensitivity, overall discriminative agreement beyond chance was weak. Variable importance analysis revealed that microbiological factors (organism and specimen source) were the strongest predictors of resistance. Conversely, spatial features (hospital ward) lacked predictive value, indicating widespread transmission. Network analysis identified a fully connected co-resistance topology (degree centrality = 4) among all sentinel antibiotics. Conclusion High-sensitivity AI models utilizing minimal metadata offer a potentially scalable triage-support approach for MDR surveillance in deeply resource-constrained health systems. Given the modest discriminatory performance, low specificity, small final analytical sample, and the absence of external validation, these findings should be interpreted as a preliminary proof-of-concept rather than evidence of a deployment-ready tool. The findings underscore critical, hospital-wide infection prevention and control (IPC) vulnerabilities and highlight the necessity of implementing digital epidemiological tracking, alongside external validation and benchmarking against alternative algorithms, to guide empiric therapy in East Africa.