Multistep validation of a post-ERCP pancreatitis prediction system integrating multimodal data: a multicenter study

Area of Science:

• Gastroenterology
• Medical Informatics
• Predictive Analytics

Background:

• Post-ERCP pancreatitis (PEP) prediction remains challenging.
• The impact of diverse data categories on PEP risk prediction is not fully understood.
• Multimodal data integration offers potential for improved PEP risk assessment.

Purpose of the Study:

• To develop and validate a multimodal prediction model for post-ERCP pancreatitis (PEP).
• To investigate the contribution of various data categories (EHR, imaging) to PEP prediction.
• To identify key risk factors for PEP using machine learning.

Main Methods:

• Retrospective collection of 1916 ERCP cases from multiple centers.
• Identification of 49 EHR features and 1 image feature related to PEP.
• Incremental model construction incorporating baseline, diagnosis, technique, and prevention features, with and without imaging data.
• Model validation using internal and external test sets and comparison of machine learning algorithms.

Main Results:

• Incorporating technique and prevention strategies significantly improved PEP prediction sensitivity and specificity compared to baseline and diagnosis features alone.
• Previous pancreatitis, NSAID use, and difficult cannulation were key EHR predictors.
• Image features demonstrated high importance in prediction models.
• A random forest model (model 8) incorporating all features achieved the best performance.

Conclusions:

• A multimodal prediction model for PEP with clinical acceptability was developed.
• Imaging and technical procedural features are critical for accurate PEP prediction.
• The findings support the use of comprehensive data integration for enhanced PEP risk stratification.