Predicting methicillin resistance in culture-confirmed invasive Staphylococcus aureus infections in the ICU: an interpretable machine learning model

Methicillin-resistant Staphylococcus aureus (MRSA) is an important cause of invasive infection in critically ill patients. Once a sterile-site culture identifies S. aureus , susceptibility testing is still required to distinguish MRSA from methicillin-susceptible S. aureus (MSSA). We developed an interpretable machine-learning model to estimate methicillin resistance during this interval using clinical data recorded within the first 24 h of ICU admission. We conducted a retrospective cohort study using MIMIC-IV. Adult ICU patients with a first qualifying culture-confirmed invasive S. aureus episode obtained more than 48 h after ICU admission were included. Six algorithms were compared using discrimination, calibration, paired AUC tests, and decision-curve analysis. SHAP quantified feature contributions to model predictions. Among 2,446 patients (1,092 MRSA [44.6%] and 1,354 MSSA [55.4%]), multistage feature selection identified 17 predictors. Random Forest achieved good discrimination (AUC 0.871, 95% CI 0.841–0.900; sensitivity 0.767; specificity 0.775; NPV 0.805). Its AUC did not differ significantly from CatBoost (ΔAUC 0.0064; DeLong P = 0.3585), LightGBM (ΔAUC 0.0075; P = 0.3802), or XGBoost (ΔAUC 0.0082; P = 0.2742). At a sensitivity-prioritizing threshold of 0.30, sensitivity was 0.932 and NPV was 0.910. SHAP identified an association pattern in which bicarbonate, lower white blood cell count, lower systolic blood pressure, age, and chronic kidney disease jointly contributed to predicted risk; these associations do not establish pathophysiological mechanisms. The model achieved good internal discrimination for methicillin resistance among patients with culture-confirmed invasive S. aureus infection. It is an auxiliary risk-stratification model for the interval between identification of S. aureus and susceptibility results, not a diagnostic or rule-out test. At the default threshold, 51/219 MRSA cases were classified as low risk; treatment decisions must therefore not rely on the model alone. Because the study used a single-centre database, independent external and prospective validation is required before implementation. Not applicable. This was a retrospective cohort study using de-identified data from a publicly available database.