Interpretable machine learning model for predicting major adverse kidney events within 30 days in critically ill patients with acute pancreatitis: development and validation in multicenter cohorts

Background Acute kidney injury is a frequent and serious complication of acute pancreatitis (AP), contributing to increased mortality and healthcare burden. Major adverse kidney events within 30 days (MAKE30) constitute a patient-centered composite endpoint for assessing kidney outcomes. However, effective tools for identifying patients at high risk of MAKE30 are lacking. This study aimed to develop and externally validate an interpretable machine learning model for predicting MAKE30 in critically ill patients with AP. Methods Three clinical datasets were used. MIMIC-IV served as the development cohort, and external validation used eICU-CRD and a retrospective cohort from the Affiliated Hospital of Southwest Medical University (AHSWMU). Feature selection combined least absolute shrinkage and selection operator (LASSO) regression, the Boruta algorithm, and multivariable logistic regression. Seven machine learning algorithms were used to develop prediction models. Model performance was assessed in terms of discrimination, calibration, and clinical utility. Shapley additive explanations (SHAP) were used to interpret and visualize the final model. Results A total of 1,574 patients with AP from multiple centers were included. Six predictors were selected for model development: serum creatinine, red cell distribution width, albumin, aspartate aminotransferase, vasopressor use, and total bilirubin. The random forest model demonstrated favorable overall performance, with an area under the receiver operating characteristic curve (AUC) of 0.863 in the internal validation cohort. Performance was consistent across the two independent external validation cohorts, with AUCs of 0.840 in eICU-CRD and 0.848 in AHSWMU. A web-based clinical risk calculator was developed to facilitate model use in clinical practice. Conclusion The model developed in this study exhibited favorable performance and robust generalizability in predicting MAKE30 among critically ill patients with AP, providing reliable digital decision support for early risk stratification and individualized intervention strategies.