Predictive models for post-ERCP pancreatitis: a systematic review and meta-analysis

TL;DR

This study reviews and evaluates existing models for predicting post-ERCP pancreatitis, finding that while they show promise, their accuracy and reliability vary, and more validation is needed for clinical use.

Contribution

A systematic review and meta-analysis of predictive models for post-ERCP pancreatitis, highlighting performance variability and the need for external validation.

Findings

Machine learning models showed higher AUC (0.84) compared to logistic regression models (0.76).

Common predictive factors included difficult cannulation, female sex, pancreatic duct dilation, and history of pancreatitis.

The pooled AUC for externally validated models was 0.79, but many models lacked external validation and had significant bias.

Abstract

Post-ERCP pancreatitis (PEP) is the most common complication following ERCP, leading to significant clinical and economic consequences. Predictive models for PEP can help identify high-risk patients and guide preventive strategies. However, the performance of these models varies, and a comprehensive evaluation is lacking. This study aims to assess the accuracy, reliability, and risk of bias in existing predictive models for PEP.

A comprehensive search was conducted across five databases (PubMed, Embase, Web of Science, Cochrane Library, and CNKI) for studies published until January 2025. Studies that developed or validated predictive models for PEP were included. Models with external validation sets were included in a meta-analysis. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and calibration. A random-effects meta-analysis was performed, with heterogeneity assessed using I² statistics. Data extraction and risk of bias were conducted using a standardized template combining the CHARMS and PROBAST tools.

Twenty-three studies (21 model development studies and 2 external validation studies) were included, presenting 21 predictive models for PEP. Nine models incorporated external validation, with one study recalibrating an existing model and another externally validating two prior models. The mean events per variable (EPV) across studies was 10.2 (2.2 to 22.4). The pooled AUC for externally validated models was 0.79 (95% CI: 0.75–0.83). Machine learning models demonstrated higher AUC (0.84) than traditional logistic regression models (0.76). Common predictive factors included difficult cannulation, female sex, pancreatic duct dilation, and a history of pancreatitis.

Predictive models for PEP show potential for improving patient risk stratification. However, variability in model performance, lack of external validation, and significant bias in many studies limit their clinical applicability. Further external validation, model refinement, and improved bias control are essential for broader clinical implementation.

https://www.crd.york.ac.uk/PROSPERO/view/CRD42024626168, identifier CRD42024626168.

Figures