Fusion prediction model for post-ERCP pancreatitis under NSAIDs prophylaxis.

BACKGROUND: Endoscopic retrograde cholangiopancreatography (ERCP) is a common procedure for treating biliary and pancreatic disorders; however, it is associated with significant complications, notably post-ERCP pancreatitis (PEP), which has an incidence ranging from 3.5 to 9.7%. OBJECTIVE: This study aims to develop machine learning-based predictive models for PEP in patients receiving prophylactic non-steroidal anti-inflammatory drugs (NSAIDs). METHODS: We retrospectively analyzed the data from 991 patients who underwent ERCP between January 2018 and February 2024. A mixed sampling technique was employed to address data imbalance. Predictive models were constructed using logistic regression, light gradient boosting machine, multi-layer perceptron, random forest, and extreme gradient boosting. The top-performing models were combined to enhance predictive accuracy. Model evaluation was conducted using a prospective cohort of 124 patients as an independent internal test set. RESULTS: Among the patients, 57 (5.75%) developed PEP. The models identified significant pre-procedure and post-procedure risk factors. The fusion model demonstrated an area under the curve of 0.84 for pre-procedure predictions and 0.86 for post-procedure predictions. High specificity and negative predictive values were noted, indicating potential clinical utility. CONCLUSION: This study presents a novel fusion model for predicting PEP in NSAID-treated patients, offering improved tools for risk stratification and clinical decision-making. Future research should focus on multi-center validation and the integration of additional predictive factors to enhance model robustness and applicability.