Explainable early prediction of short-term severe deterioration in gallstone-related biliary disease using MIMIC-IV, eICU, and local real-world data

Patients with gallstone-related biliary disease may deteriorate rapidly before definitive source control. This study aimed to develop an interpretable 0-24-hour prediction model for short-term severe deterioration in gallstone-related biliary disease and validate it across eICU-CRD and a local real-world cohort. MIMIC-IV was used for model development, eICU-CRD for primary external validation, and local clinical data for secondary validation. Eligible patients had gallstone disease, cholecystitis, cholangitis, or biliary obstruction. The outcome was a composite of in-hospital or ICU death, vasopressor use, mechanical ventilation, continuous renal replacement therapy, or prolonged ICU stay. Feature selection, imputation, scaling, model selection, and threshold determination were restricted to MIMIC-IV development data. Performance was assessed using AUROC, AUPRC, Brier score, calibration, decision curves, threshold consequences, subgroup analyses, bootstrap risk uncertainty, and SHAP explanations. The study included 2818 MIMIC-IV stays, 1492 eICU-CRD stays, and 115 local patients, with event rates of 58.4%, 65.7%, and 52.2%, respectively. The locked elastic net model retained 12 routine early predictors. AUROC was 0.70 in MIMIC-IV internal validation, 0.64 in eICU-CRD external validation, and 0.63 in local validation. Calibration, net benefit, threshold consequences, and SHAP analyses supported use as a bedside risk-support tool rather than a replacement for clinical judgment. A routine early model captured multi-system deterioration signals in gallstone-related biliary disease and showed transportable, moderate discrimination across external cohorts.