A machine learning model for in-hospital mortality prediction in patients with sepsis-related myocardial injury

Sepsis-related myocardial injury (SRMI) is associated with high mortality, but the transportability of prediction models across heterogeneous cohorts remains uncertain. We developed and externally validated a machine learning model for in-hospital mortality prediction in SRMI and evaluated its discrimination, calibration, interpretability, and distributional shift across cohorts. Using MIMIC-IV for development and MIMIC-III and a local cohort for external validation (n = 1,226, 962, and 280, respectively), we compared eight classification algorithms. A two-stage feature-selection strategy, recursive feature accumulation and sequential elimination with incremental feature evaluation, ranked 31 candidate variables to derive a parsimonious predictor set. The selected model was evaluated for discrimination, precision-recall performance, calibration, clinical utility, and risk stratification. Logistic recalibration, population stability index (PSI) analysis, and SHapley Additive exPlanations (SHAP) were used to assess probability alignment, cross-cohort distributional shift, and model interpretability. Feature selection reduced 31 candidate variables to 8 predictors with little loss of cross-validated performance (AUC, 0.819 with 31 variables versus 0.810 with 8 variables). The selected random forest model achieved an AUC of 0.819 (bootstrap 95% CI, 0.787-0.848) and a PR-AUC of 0.702 (0.657-0.751) in MIMIC-III, and an AUC of 0.758 (0.693-0.816) and a PR-AUC of 0.525 (0.435-0.630) in the local cohort. Recalibration reduced the Brier score from 0.162 to 0.147 in MIMIC-III and from 0.233 to 0.196 in the local cohort. PSI indicated minimal distributional shift between MIMIC-IV and MIMIC-III but substantial shift in the local cohort, particularly for SOFA score, respiratory rate, and prothrombin time. SHAP identified prothrombin time, invasive ventilation, lactate, temperature, and urine output as major contributors to model predictions. In two external cohorts, an 8-variable random forest model showed moderate discrimination for mortality prediction in SRMI. The evaluation of external performance, calibration, distributional shift, and SHAP-based explanations provides an assessment of model transportability. Prospective validation is needed before clinical deployment.Author summary: Sepsis can damage the heart and is associated with a high risk of death in intensive care. Clinicians need practical ways to identify patients with sepsis-related myocardial injury who may need closer monitoring after the first day in the intensive care unit. We developed a prediction model using routinely collected information from the first 24 hours of ICU care, including age, illness severity, breathing support, urine output, vital signs, lactate, and prothrombin time. We developed the model in one large critical-care database and tested it in a second database and an independent local hospital cohort. The model showed moderate ability to distinguish survivors from non-survivors in both external cohorts. We also assessed whether its predicted risks were well calibrated, how patient characteristics differed between hospitals, and which variables contributed most to predictions. The model is intended to support early risk stratification, not to replace clinical judgment or established severity scores. Because the local cohort was small and the study was retrospective, prospective multicenter validation and local assessment are needed before clinical use.