Machine learning for the prediction of post-ERCP pancreatitis risk: A proof-of-concept study

Area of Science:

• Medical informatics
• Gastroenterology
• Machine learning in healthcare

Background:

• Predicting Post-Endoscopic Retrograde Cholangiopancreatography (ERCP) pancreatitis (PEP) is crucial for patient management.
• Previous studies using standard statistical methods have limited accuracy in identifying PEP risk factors.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting PEP probability.
• To identify significant clinical features associated with PEP risk.

Main Methods:

• An international, multicenter, prospective cohort of ERCP patients was analyzed.
• Gradient boosting (GB) and logistic regression (LR) models were trained and tested.
• SHAP (Shapley Additive exPlanations) was used to interpret model features.

Main Results:

• The GB model demonstrated superior performance with an Area Under Curve (AUC) of 0.671 in the test set, outperforming LR.
• Key predictors for PEP included bilirubin levels, age, BMI, procedure duration, and previous procedures.
• Several pre-procedural factors were identified as significant contributors to PEP risk.

Conclusions:

• Gradient boosting models significantly outperform logistic regression for PEP prediction.
• ML models can identify novel and existing clinical features relevant to PEP risk.
• These findings can aid in refining patient selection and procedural strategies to mitigate PEP.