Development and Validation of a Practical Prediction Model for Post-ERCP Pancreatitis using Machine Learning

Tianyu De¹Guohui Du²Hongkun Yin²Hao Wang^2*Wei Wang¹Tian Ma¹Junbai Ma³Qi Wang^1*Hao Wang^4*

• ¹General Hospital of Ningxia Medical University, Yinchuan, China
• ²School of Clinical Medicine, General Hospital of Ningxia Medical University, Yinchuan, China
• ³School of Basic Medical Sciences, Ningxia Medical University, Yinchuan, Chile
• ⁴Department of Pathogenic Biology and Medical Immunology, School of Basic Medical Sciences, Ningxia Medical University, Yinchuan, China

Background: Post-endoscopic retrograde cholangiopancreatography (ERCP) pancreatitis (PEP) is one most frequent and severe complication of ERCP. In consideration of recent advancements in both endoscopic and artificial intelligence research, it is possible to construct a practical risk prediction model to facilitate the identification of PEP patients at elevated risk. Aim: We developed and validated a concise predictive model for post-ERCP pancreatitis risk with logistic regression (LR), LightGBM, Support Vector Machine (SVM), XGBoost, and Multilayer Perceptron (MLP) neural network models. Methods: We selected 688 patients undergone ERCP to form the basic dataset, with 70% for training and 30% for validation. Subsequently, Stepwise Backward Selection Based on Logistic Regression was utilized to select pertinent clinical features, incorporating the machine learning (ML) models to construct the final predictive model. The efficacy of the model was evaluated by various metrics. These newly identified clinical features were then incorporated into a simplified, points-based risk scoring system for potential bedside application and further evaluation. Results: Based on the collected data and the results of stepwise backward regression, we identified the following features as potentially significant clinical variables that influence the risk of post-ERCP pancreatitis: periampullary diverticulum, pancreatic stent placement, pancreatic guidewire passages, dilation of the extrahepatic bile duct, age, and coronary artery disease, and constructed a prediction model. Following this, several ML models were constructed to assess the performance of this model. All ML models demonstrated superior performance to conventional logistic regression (LR) models in terms of AUC curves, with XGBoost, SVM, LightGBM, and MLP models all achieving at least acceptable performance levels. Finally, we developed a simplified scoring system based on LightGBM model with an AUC of 0.75. Conclusions: We developed and validated a concise predictive model for post-ERCP pancreatitis risk, and a simplified scoring system based on the LightGBM model. This model facilitates individual risk prediction and preventive strategy selection.