Machine Learning-Based Prediction of Short-Term Mortality in Patients With Severe Acute Pancreatitis: A Multicenter Retrospective Cohort Study

Objective: To identify risk factors for short-term mortality in severe acute pancreatitis (SAP), establish a predictive model for early high-risk patient identification, and guide clinical decision-making. Methods: SAP patients admitted to the Affiliated Hospital of Xuzhou Medical University from September 2018 to September 2025 were enrolled, divided into mortality and survival groups by 28-day prognosis. Clinical data were collected. Features were strictly selected through Least Absolute Shrinkage and Selection Operator (LASSO) regression, Boruta algorithm, and Recursive Feature Elimination (RFE). Seven machine learning (ML) models were built, with external validation using Medical Information Mart for Intensive Care IV (MIMIC-IV) data. Model performance was evaluated through receiver operating characteristic (ROC) curves, calibration curves, and decision curves. SHapley Additive exPlanations (SHAP) analysis was used to interpret contributions of important features, and a web-based calculator was developed for visualization. Results: Ten features were selected. The Gradient Boosting Machine (GBM) model had the best generalization, with area under the ROC curve (AUC) values of 0.964 (95% CI: 0.942-0.987, training), 0.927 (95% CI: 0.885-0.970, testing), and 0.811 (95% CI: 0.772-0.851, validation). Calibration curves confirmed predicted-actual consistency; decision curves showed net clinical benefit. SHAP analysis identified the ranking of feature importance as follows: mechanical ventilation, age, blood urea nitrogen, urine output, lactate, total bilirubin, platelet count, congestive heart failure history, red blood cell distribution width, and serum creatinine. The web-based calculator had good clinical applicability. Conclusion: The GBM model demonstrates the best performance in predicting short-term mortality in SAP patients.