An interpretable machine learning model for predicting the risk of post-endoscopic retrograde cholangiopancreatography complications in elderly patients with choledocholithiasis

Background: Identifying an effective method to predict the risk of post-endoscopic retrograde cholangiopancreatography (ERCP) complications is of significant importance for developing personalized treatment plans and improving outcomes. Previous studies have notable limitations; hence, we have developed and validated an interpretable machine learning model for predicting complications. Methods: We collected data from patients over 60 years old with choledocholithiasis treated at the First Medical Center of Chinese PLA General Hospital between January 2010 and December 2024. We constructed models using nine machine learning methods, including random forest (RF). The predictive performance of the models was compared using evaluation metrics, such as the area under the receiver operating characteristic curve (AUC). The SHapley Additive exPlanations (SHAP) method was used to rank the importance of features and to interpret the final model. Results: Among the nine models, the RF model performed the best. The RF model was able to accurately predict the postoperative complication risk of patients in both the internal validation of the training set [AUC = 0.970 ± 0.021, 95% CI = (0.956, 0.991)] and the test set [AUC = 0.946 ± 0.010, 95% CI = (0.926, 0.996)]. Calibration curves showed a high degree of concordance between predicted and observed risks. The final RF model included nine variables: distal common bile duct angle, cannulation time, INR, number of difficult stone factors, CRP, age, stone count, stent, and length of the distal common bile duct. Conclusion: We have developed and validated an interpretable RF model for predicting the risk of post-ERCP complications in elderly patients with choledocholithiasis, and proposed targeted preventive strategies.