Data Challenge: AI-Driven Clinical Decision Support for Serious Bacterial Infections Using Global Antimicrobial Resistance (AMR) Surveillance Data

Our project will transform large-scale, retrospective AMR surveillance data into a novel, dual-purpose AI-driven clinical decision support system for clinical care in resource-limited settings. This work builds upon our previously developed AI model for predicting Serious Bacterial Infection (SBI), which was trained and internally validated on 11,466 records from Northern Nigeria. As concluded in our initial study, the model now requires external validation and calibration on diverse datasets to ensure global transportability and clinical utility. The central objectives are twofold. Firstly, we will leverage the extensive geographic and demographic coverage of selected Vivli AMR Register datasets to externally validate, retrain, and enhance the generalizability of our existing SBI prediction model. Secondly, we will develop and validate a “de novo” machine learning model to predict AMR at the point of care, using the rich Minimum Inhibitory Concentration, demographic, and microbiological data provided, which will be embedded into the AI system. This innovative approach addresses a critical need by converting surveillance data into an actionable tool for guiding empirical antibiotic selection, especially in low-resource settings. Our methodology will employ an ensemble of machine learning models (XGBoost, Random Forest, logistic regression) using R and Python, with rigorous performance evaluation using sensitivity, specificity, AUROC, and balanced accuracy, adhering to TRIPOD statement principles. Model explainability will be assessed using SHAP. The anticipated result is a validated framework for an integrated tool that addresses two critical questions: SBI risk and likely pathogen resistance. This empowers frontline healthcare workers, optimises patient triage, promotes antibiotic stewardship, and ultimately improves clinical outcomes worldwide. All analytical code will be shared via GitHub to ensure reproducibility.