Abstract
BACKGROUND: Post-endoscopic retrograde cholangiopancreatography pancreatitis (PEP) is the most common and clinically significant complication of ERCP, with an incidence of 3.5-9.7% in general populations and up to 14.7% in high-risk groups, leading to considerable morbidity, mortality, and healthcare costs. Although numerous multivariable prediction models have been developed, their predictor sets, methodological rigor, and clinical applicability remain highly variable. METHOD: We conducted a PRISMA 2020-compliant systematic review and meta-analysis, prospectively registered in PROSPERO (CRD42024556967). Nine databases were searched to June 1, 2024, for studies developing or validating multivariable PEP risk prediction models. Data on study/model characteristics, predictors, and performance metrics were extracted. Risk of bias was assessed with PROBAST, and study quality with the Newcastle-Ottawa Scale. Random-effects meta-analyses pooled (i) PEP incidence, (ii) associations of individual predictors, and (iii) overall model performance. RESULTS: Twenty-four studies (26 models; n = 38,016) published from 2002-2024 were included, predominantly retrospective cohorts from East Asia (n = 16). The pooled PEP incidence was 8.48% (95% CI: 6.90-10.39%; I² = 96.4%), highest in East Asia and retrospective cohorts. Strongest predictors included pancreatic duct cannulation (OR=3.50), pancreatic injection (OR=3.50), previous pancreatitis (OR=3.32), and pancreatic guidewire use (OR=2.63); additional consistent factors were female sex, difficult cannulation, elevated bilirubin, low albumin, choledocholithiasis, and prolonged procedure time. The pooled odds ratio for model performance was 0.81 (95% CI: 0.78-0.84; I² = 83.5%), with AUCs ranging 0.560-0.915, though calibration was infrequently reported (38%) and external validation undertaken in only 46%. PROBAST indicated high overall risk of bias, chiefly in the analysis (92%) and participants (100%) domains. CONCLUSION: Current PEP prediction models generally demonstrate moderate-to-high discrimination but are limited by suboptimal calibration, inadequate external validation, and methodological heterogeneity. Future research should adhere to TRIPOD guidelines, employ multicenter large-sample designs, retain continuous predictors, address missing data with robust imputation methods, and conduct comprehensive temporal, geographic, and domain-specific validation. Integration of artificial intelligence/machine learning with conventional modeling and embedding validated tools into clinical workflows may enhance predictive accuracy and real-world utility.