Machine learning-derived predictive model for post-ERCP pancreatitis in patients with common bile duct stones: a retrospective multicenter study.

Background: Common bile duct stones (CBDS) are the primary indication for endoscopic retrograde cholangiopancreatography (ERCP), yet post-ERCP pancreatitis (PEP) remains a significant complication due to its multifactorial etiology. This study aimed to identify core predictors and develop an optimized predictive model for PEP.

Methods: We retrospectively enrolled patients who underwent ERCP in three centers between March 2019 and March 2024. Potential predictors and their importance were evaluated with four machine learning (ML) algorithms. Predictive models were developed using logistic regression and assessed for discrimination, calibration, and clinical utility.

Results: A total of 1758 patients were included in the training (n = 917), testing (n = 392), validation 1 (n = 366), and validation 2 (n = 83) cohorts. The incidences of PEP were 6.7%, 6.6%, 10.1%, and 12.0%, respectively, with no significant difference among them (p = 0.063). Using ML, eight critical predictors were identified: age, direct bilirubin, serum calcium, γGT, cannulation attempts, transpancreatic precut, pancreatic guidewire passage, and endoscopic papillary balloon dilation (EPBD) duration. Model 3, incorporating serum calcium (OR: 2.50, p = 0.002), transpancreatic precut (OR: 4.61, p < 0.001), pancreatic guidewire passage (OR: 3.62, p < 0.001), and EPBD duration (OR: 2.25, p = 0.009), exhibited the highest AUC (0.845) and superior sensitivity (83.2%). Internal and external validations confirmed robustness and generalizability of the model, demonstrating excellent predictive performance and clinical utility.

Conclusion: We established and validated an optimized predictive model for PEP using four key predictors, enhancing early identification and intervention after ERCP for patients with CBDS.