Predicting Post-ERCP Pancreatitis Using Machine Learning: Risk Stratification and Feature Importance Analysis

Area of Science:

• Gastroenterology
• Medical Informatics
• Predictive Analytics

Background:

• Post-endoscopic retrograde cholangiopancreatography pancreatitis (PEP) remains unpredictable despite risk factor understanding.
• Developing predictive models for PEP is crucial for patient safety.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) model for predicting the risk of PEP.
• To identify key predictors of PEP in patients undergoing ERCP.

Main Methods:

• Trained CatBoost and eXtreme Gradient Boosting algorithms on prospective ERCP registry data (2022-2024).
• Assessed model performance using area under the receiver operating characteristic (AUC) with 10-fold cross-validation.
• Included patients with naïve papilla undergoing ERCP.

Main Results:

• The CatBoost model, using eight features (age, sex, papilla morphology, cannulation details, bilirubin, CBD diameter, stone extraction), achieved an AUC of 68.8%.
• The model demonstrated good sensitivity (70.4%) and specificity (67.2%), with a negative predictive value of 92.0%.
• PEP incidence varied significantly across risk groups: 5.7% (low), 21.0% (intermediate), and 40.0% (high).

Conclusions:

• Machine learning shows significant promise for predicting PEP risk.
• Future research should explore multicenter data, multimodal data integration, severity stratification, and real-time applications.