Abstract
Post-endoscopic retrograde cholangiopancreatography (ERCP) pancreatitis (PEP) is a common complication in patients undergoing ERCP for choledocholithiasis, yet effective predictive models are lacking. This study included 2,247 patients who underwent ERCP for complete stone removal at the First Affiliated Hospital of USTC from January 2015 to January 2023. Six machine learning algorithms were utilized, incorporating 25 clinical parameters, to develop a predictive model for PEP risk. The random forest (RF) algorithm achieved the highest accuracy, with an area under the receiver operating characteristic curve (AUC) of 0.947 in the internal dataset. Key risk factors for PEP identified include difficult cannulation, a history of pancreatitis, smaller common bile duct diameter, and female gender. Validation with datasets from 12 external centers showed AUC values ranging from 0.576 to 0.913, with an average of 0.768. An interactive R Shiny web application was also developed, offering a user-friendly tool for predicting PEP risk and enabling individualized management.