AI-Based Prediction of Post-ERCP Pancreatitis: A Comparative Study Using Tabular, Image, and Multimodal Data.

Area of Science:

• Gastroenterology and Endoscopy
• Medical Informatics and Machine Learning
• Clinical Prediction Modeling

Background:

• Post-Endoscopic Retrograde Cholangiopancreatography Pancreatitis (PEP) is a significant complication of ERCP, affecting 2-10% of patients.
• Early prediction of PEP risk is vital for timely intervention and improved patient outcomes.
• This study evaluated the predictive capabilities of clinical data, endoscopic images, and multimodal approaches for PEP.

Purpose of the Study:

• To comparatively assess the performance of machine learning models using tabular clinical data, endoscopic images, and multimodal fusion for predicting PEP.
• To identify key clinical and imaging features associated with PEP risk.
• To explore the potential of advanced AI techniques in managing ERCP complications.

Main Methods:

• Retrospective analysis of clinical and endoscopic data from a single center.
• Application of XGBoost for tabular data and EfficientNet-B0, ResNet50, DenseNet201 for image data.
• Development of a multimodal contrastive learning framework combining image and tabular features.
• Implementation of class imbalance mitigation techniques and performance evaluation using AUC, sensitivity, and F1-score.

Main Results:

• The XGBoost model using tabular clinical data achieved the highest predictive performance (AUC 0.95).
• Image-based models, particularly ResNet50, showed moderate performance (AUC 0.76).
• The multimodal model demonstrated lower predictive power (AUC 0.57). Key predictors included cannulation time, ampulla type, and patient age.

Conclusions:

• Structured clinical data currently offer the most robust predictive signals for PEP compared to endoscopic images or multimodal fusion in this cohort.
• Class imbalance due to the low incidence of PEP remains a challenge.
• Future multicenter studies with larger datasets and advanced techniques are needed to enhance predictive accuracy and clinical utility.