Abstract / Summary
Abstract Background Patients with both heart failure (HF) and type 2 diabetes mellitus (T2DM) are susceptible to sepsis, but no validated tool addresses early risk stratification in this population. We developed and validated interpretable machine learning models using two critical care databases. Methods Adult intensive care unit patients with HF and T2DM were identified in MIMIC-IV (2008–2022); an eICU cohort was used for external validation. Feature selection combined univariate and multivariable analyses, least absolute shrinkage and selection operator regression, and the Boruta algorithm. Seven models were tuned by grid search with five-fold cross-validation. Discrimination, calibration, clinical utility, and interpretability were assessed with the area under the receiver operating characteristic curve (AUC-ROC), Brier score, decision curve analysis, and SHapley Additive exPlanations (SHAP), respectively. Results The MIMIC-IV cohort comprised 7,806 patients, of whom 4,195 (53.74%) had sepsis; the external eICU cohort comprised 633 patients. Seven predictors were retained: acute kidney injury, mechanical ventilation, white blood cell count, body temperature, Glasgow Coma Scale score, anion gap, and blood lactate. XGBoost had the highest overall performance, with AUC-ROC values of 0.712 (95% CI: 0.691–0.733) in internal validation and 0.658 (95% CI: 0.551–0.756) in external validation. SHAP ranked acute kidney injury, Glasgow Coma Scale score, white blood cell count, mechanical ventilation, and body temperature as the leading contributors. Conclusions The XGBoost model provided moderate discrimination with interpretable predictions based on routinely available clinical variables. It may support early sepsis risk assessment in intensive care patients with HF and T2DM, but prospective multicenter validation is needed before clinical use.