An interpretable machine learning model for early prediction of ESBL-producing bacteraemia in the emergency department

Empirical antibiotic therapy in the emergency department (ED) is often initiated before susceptibility results are available, risking treatment failure in bloodstream infections caused by extended-spectrum β-lactamase (ESBL)-producing Escherichia coli and Klebsiella spp. We retrospectively analysed 3,138 adult sepsis episodes (training: n = 1,863, ESBL + = 419, ESBL− = 1,444; test: n = 1,275, ESBL + = 373, ESBL− = 902) treated at Inha University Hospital between 2013 and 2022 and trained a machine-learning model to predict ESBL-producing organisms using routinely available clinical and laboratory data. We used stratified 5-fold cross-validation and computed SHapley Additive exPlanations (SHAP) values within folds to avoid information leakage. We performed SHAP-guided stepwise feature elimination based on training performance. The final model, trained on the full training cohort (2013–2019), achieved an AUROC of 0.78 on an independent test set (2020–2022). This interpretable pipeline may support empirical antibiotic selection and antimicrobial stewardship in the ED.