https://orcid.org/0009-0005-7925-7347
Figures
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.
Citation: Mao Y, Liu Q, Fan H, He W, Zhang C, Ouyang X, et al. (2025) Risk prediction model for post-endoscopic retrograde cholangiopancreatography pancreatitis: A systematic review and meta-analysis. PLoS One 20(9): e0332378. https://doi.org/10.1371/journal.pone.0332378
Editor: Suphakarn Techapongsatorn, Faculty of Medicine Vajira Hospital, Navamindradhiraj University, THAILAND
Received: November 5, 2024; Accepted: August 29, 2025; Published: September 15, 2025
Copyright: © 2025 Mao et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the manuscript and its Supporting Information files.
Funding: This research project was supported by Xianyang Science and Technology Planning Project (Grant Number: L2023-ZDYF-SF-055). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Endoscopic retrograde cholangiopancreatography (ERCP) is widely utilized for diagnosing and treating biliopancreatic diseases [1,2]. Compared with traditional surgery, ERCP offers several advantages, including reduced surgical trauma, faster postoperative recovery, shorter treatment durations, and briefer hospital stays [3]. However, as an invasive diagnostic and therapeutic procedure, ERCP carries inherent risks, particularly procedure-related complications. Among these, post-ERCP pancreatitis (PEP) is the most common, with an incidence of 3.5–9.7% in the general population and up to 14.7% in high-risk patients. The mortality rate associated with PEP ranges from 0.1% to 0.7% [2,4,5]. While mild cases may simply prolong hospitalization, severe PEP can result in pancreatic necrosis, multiple organ failure, or even death [6,7]. These adverse outcomes not only compromise procedural success and patient prognosis but also impose substantial economic burdens on patients, healthcare systems, and society.
To mitigate these consequences, accurate identification of patients at high risk of PEP is essential. Risk prediction models offer clinicians a tool to estimate individual PEP risk and tailor preventive strategies accordingly. By stratifying patients according to predicted risk, unnecessary admissions can be reduced, preventive measures can be targeted, and overall healthcare costs can be lowered-ultimately improving patient quality of life.
In recent years, multiple PEP risk prediction models have been developed using different predictors, statistical methods, and validation strategies. However, their methodological quality, external validity, and clinical applicability vary considerably. Moreover, the predictors incorporated into these models differ substantially across studies, reflecting variations in patient populations, procedural techniques, and study designs. While such heterogeneity is inevitable, it raises important questions: which predictors show consistent, clinically meaningful associations with PEP? How well do existing models perform overall? And what is the true incidence of PEP in the populations studied?
A systematic review and meta-analysis provides a rigorous approach to address these questions. By synthesizing available evidence, such an approach can: (i) quantify the pooled incidence of PEP across diverse populations; (ii) identify predictors with consistent, statistically significant associations with PEP; (iii) evaluate the overall predictive performance of existing models, considering both discrimination and calibration; (iv) assess the methodological quality and risk of bias of the included studies.
Therefore, the aim of this study was to systematically review and meta-analyze existing PEP risk prediction models, with a focus on their predictive performance, methodological rigor, and clinical applicability. Through this synthesis, we aim to provide robust, evidence-based recommendations to guide both clinical decision-making and future model development.
Methods
This study was registered in PROSPERO (registration number: CRD42024556967) and conducted in accordance with the PRISMA 2020 guidelines. For detailed information, please refer to S1 Table.
Search strategy
We systematically searched PubMed, Web of Science, The Cochrane Library, Embase, Cumulative Index to Nursing and Allied Health Literature (CINAHL), China National Knowledge Infrastructure (CNKI), Wanfang Database, China Science and Technology Journal Database (VIP), and SinoMed from inception to June 1, 2024. The detailed search strategies for each database are provided in S2 and S3 Tables.
Inclusion and exclusion criteria
The inclusion criteria were as follows: (i) studies involving patients underwent ERCP; (ii) observational study design (cohort or case-control), or interventional studies that developed or validated a PEP prediction model; (iii) reported outcome of post-ERCP pancreatitis (PEP); and (iv) inclusion of a multivariable prediction model. The exclusion criteria were as follows: (i) studies that only assessed risk factors without constructing a prediction model; (ii) studies without available full text; (iii) gray literature, including conference abstracts and agency reports; (iv) duplicate publications; and (v) studies not written in English or Chinese.
Study selection and screening
Two reviewers independently screened titles/abstracts and full texts according to the inclusion criteria. Disagreements were resolved by a third reviewer.
Data extraction
Data were extracted into a standardized form and categorized as: (i) study characteristics: first author, year, country/region, research design, data source, study period, and PEP diagnostic criteria; (ii) model characteristics: sample size, outcome event rate, events per variable (EPV), model development method, variable selection method, handling of missing data, and treatment of continuous variables; (iii) model performance: discrimination (e.g., AUC/C-statistic), calibration, type of validation (internal vs. external), and model presentation format; and (iv) predictors: number and type of candidate variables and final predictors included in the model.
Quality assessment
The methodological quality and applicability were assessed using the Prediction model Risk Of Bias ASsessment Tool (PROBAST) [8]. The Newcastle-Ottawa Scale (NOS) was employed to assess the quality of observational studies [9]. Two reviewers (YM. & QL.) independently rated the certainty of evidence using the GRADE system (Grading of Recommendations Assessment, Development and Evaluation) [10], with disagreements resolved through discussion.
Data synthesis and statistic analysis
We performed three complementary meta-analyses:
- 1. Pooled incidence of PEP
Extracted incidence data (events/total) from each study. Proportions were pooled using a random-effects model with logit transformation to stabilize variances. Heterogeneity was assessed with Cochran’s Q test and quantified by I²; thresholds of 25%, 50%, and 75% indicated low, moderate, and high heterogeneity, respectively. Subgroup analyses were planned by geographic region, study design, diagnostic definition and validation method.
- 2. Pooled associations of individual predictors
For predictors reported in ≥3 studies, effect estimates (odds ratios [OR], relative risks [RR], hazard ratios [HR]) and 95% CIs were extracted; RRs and HRs were converted to ORs if necessary. Random-effects meta-analyses (restricted maximum likelihood estimator) were performed separately for each predictor. Heterogeneity and publication bias (Egger’s test, funnel plots) were evaluated when ≥10 studies were available. Sensitivity analyses excluded high risk-of-bias studies or those with small sample sizes (<100 participants).
- 3. Pooled predictive performance of models
Discrimination was quantified by the AUC (C-statistic). AUC values were pooled on the logit scale and back-transformed for presentation. Priority was given to externally validated results; internal validation was used if external validation was unavailable.
All analyses were performed using R (version 4.4.1; R Foundation for Statistical Computing, Vienna, Austria) with the meta, metafor, and mada packages. Two-tailed p-values <0.05 were considered statistically significant.
Results
Study selection
Our search identified 962 records, with 57 duplicates removed. After screening 905 articles, 783 were excluded as irrelevant. An further 98 full-text articles were excluded for the following reasons: conference abstracts (n = 26), no risk prediction model (n = 36), fewer than two predictors (n = 2), abstract only (n = 23), and not primary literature (n = 11). In total, 24 studies [11–34] reporting 26 PEP prediction models met the inclusion criteria (Fig 1).
Study characteristics
The 24 studies were published between 2002 and 2024, with 20 published in the last five years. Sixteen studies were conducted in China, four in Japan, two in the United States, one in Korea, and one in Italy. Study designs comprised 20 retrospective cohort, two prospective cohort, one cross-sectional study, and one case-control study. Diagnostic criteria for PEP varied: 13 studies used the Cotton Expert Consensus Criteria [35], nine the Atlanta Criteria [36], one the Chinese MDT Consensus [37], and one the ASGE criteria [38]. Sample sizes ranged from 312 to 6,731. The basic characteristics of the included studies are presented in Table 1.
Model development methods included logistic regression (n = 21), propensity score analysis (n = 1), gradient boosting (n = 1), and random forest (n = 1). Candidate predictors ranged from 3 to 22, the most frequent were pancreatic duct cannulation (n = 11), difficult cannulation (n = 10), pancreatic injection (n = 9), and female sex (n = 9) (Table 2 and Fig 2).
Pooled incidence of PEP
Incidence data were available for 24 studies (n = 38,016 patients). The pooled incidence was 8.48% (95% CI: 6.90–10.39%) with high heterogeneity (I² = 96.4%) (Fig 3 and S1 Fig).
This forest plot presents pooled incidence estimates (%) and 95% confidence intervals (CIs) derived from random-effects meta-analyses using the inverse variance method and logit transformation. Subgroups are arranged into four thematic categories: region, study design, outcome definition, and validation method. The vertical dashed line represents the overall pooled incidence across all included studies (n = 37; 38,016 patients), estimated at 8.48% (95% CI: 6.90–10.39%). Point estimates are indicated by circles, with horizontal lines showing the 95% CIs. Numerical values to the right of each line represent the pooled proportion and its 95% CI. Heterogeneity within subgroups was quantified using the I² statistic, and between-group differences were evaluated using the Q test. The highest incidence was observed in East Asia and in studies with external validation only, whereas the lowest incidence occurred in North America and case–control study designs.
By geographic region.
Subgroup analysis demonstrated significant geographical variation in PEP incidence (p = 0.0006). The highest incidence was observed in East Asia (9.27%, 95% CI: 7.40–11.55%; I2= 96.5%), followed by Europe (6.09%, 95% CI: 4.84–7.62%; single study), North America (4.69%, 95% CI: 3.40–6.45%; I2= 86.4%), and other regions (4.74%, 95% CI: 3.40–6.57%; I2= 51.1%) (S2 Fig).
By study design.
When stratified by study design, retrospective cohort studies reported the highest incidence (9.05%, 95% CI: 7.24–11.25%; I2 = 96.4%), followed by prospective cohort studies (5.78%, 95% CI: 4.98–6.70%; I2 = 0%), cross-sectional studies (5.22%, 95% CI: 3.96–6.86%; single study), and case–control studies (4.02%, 95% CI: 3.52–4.58%; single study). The difference between subgroups was statistically significant (p < 0.0001) (S3 Fig).
By diagnostic definition.
The pooled incidence was 9.26% (95% CI: 6.25–13.51%; I2 = 96.8%) in studies using the Atlanta definition, 10.23% (95% CI: 8.52–12.24%; single study) in those adopting the Chinese definition, and 7.96% (95% CI: 6.14–10.27%; I2= 96.6%) in studies applying other definitions. No statistically significant difference was observed between subgroups (p = 0.2919) (S4 Fig).
By validation method.
Marked variation was also detected when stratified by validation approach (p < 0.0001). Studies with external validation only reported the highest incidence (27.12%, 95% CI: 23.81–30.70%; I2 = 0%), followed by those with both internal and external validation (10.63%, 95% CI: 7.96–14.08%; I2 = 83.8%), no validation (8.55%, 95% CI: 4.92–14.46%; I2 = 98.1%), and internal validation only (6.75%, 95% CI: 5.61–8.10%; I2 = 91.8%) (S5 Fig).
Meta-analysis of individual predictors
The pooled effects of 11 risk factors for PEP are presented in Table 3 and S6 Fig, including odds ratios (ORs), 95% confidence intervals (CIs), and heterogeneity metrics (I2 and Tau2). Subgroup analyses by region (East Asia vs. North America) were performed where data permitted.
High-risk factors with moderate heterogeneity.
Pancreatic duct cannulation (OR = 3.50, 95% CI: 1.93–6.34) and pancreatic injection (OR = 3.50, 95% CI: 1.71–7.14) showed the strongest associations, though with substantial heterogeneity (I2 = 88% for both). Difficult cannulation had a higher effect size in North America (OR = 3.44, 95% CI: 1.34–8.81, I2 = 0%) compared to East Asia (OR = 2.56, 95% CI: 1.54–4.27, I2 = 77%).
Factors with low heterogeneity.
Previous pancreatitis (OR = 3.32, 95% CI: 2.37–4.64, I2 = 25%) and PGW (OR = 2.63, 95% CI: 2.02–3.43, I2 = 6%) were statistically significant with minimal heterogeneity.
Non-significant factors.
Age, TBIL, calculus of common bile duct, and operation time had wide CIs crossing 1, despite high heterogeneity (I2 > 70%).
Heterogeneity and bias assessment.
High heterogeneity (I2 > 75%) was observed for pancreatic duct cannulation, pancreatic injection, and female gender, possibly due to procedural differences across studies. Publication bias (Egger’s test) was significant only for pancreatic duct cannulation (p = 0.039), warranting caution in interpretation.
Predictive performance of models
Our systematic review and meta-analysis included 13 studies evaluating risk prediction models for ERCP pancreatitis. The pooled analysis demonstrated that these prediction models had significant predictive value, with an overall odds ratio (OR) of 0.80 (95% CI: 0.77–0.84) using a random-effects model. Considerable heterogeneity was observed among studies (I2 = 82.0%, τ² = 0.0041, p < 0.0001), suggesting substantial variation in model performance across different settings (Fig 4).
Subgroup analyses
Dataset type.
Analysis stratified by dataset type (development vs. validation groups) revealed no significant difference between subgroups (χ²₁ = 1.44, p = 0.2303). The development group showed an OR of 0.78 (95% CI: 0.72–0.85, I2 = 89.8%), while the validation group demonstrated an OR of 0.82 (95% CI: 0.78–0.87, I2 = 47.3%) (S7A Fig).
Diagnostic criteria.
Subgroup analysis by diagnostic criteria (Atlanta classification, Chinese consensus, and Cotton criterion) showed no statistically significant differences between subgroups (χ²₂ = 4.57, p = 0.1020). Models using Atlanta classification had an OR of 0.81 (95% CI: 0.76–0.86, I2 = 67.2%), while those using Cotton criterion showed an OR of 0.77 (95% CI: 0.70–0.85, I2 = 71.9%) (S7B Fig).
Study design.
Significant differences were observed across study designs (χ²₃ = 45.21, p < 0.0001). Retrospective cohort studies (OR = 0.82, 95% CI: 0.80–0.85, I2 = 34.3%) constituted the majority of included studies. Cross-sectional studies showed higher predictive values (OR = 0.84, 95% CI: 0.70–1.00, single study), while prospective cohort studies demonstrated lower values (OR = 0.70, 95% CI: 0.64–0.76) (S7C Fig).
Validation method.
Models with both internal and external validation showed the most consistent results (OR = 0.86, 95% CI: 0.85–0.88, I2 = 0%), differing significantly from those with internal validation only (OR = 0.79, 95% CI: 0.72–0.85, I2 = 72.5%) or no validation (OR = 0.80, 95% CI: 0.73–0.87, I2 = 90.5%) (χ²₂ = 14.64, p = 0.0007) (S7D Fig).
Geographic region.
Significant regional variations were observed (χ²₂ = 46.69, p < 0.0001). East Asian studies predominated (OR = 0.82, 95% CI: 0.80–0.85, I2 = 27.2%), while European (OR = 0.70, 95% CI: 0.64–0.76) and North American studies (OR = 0.71, 95% CI: 0.68–0.74) showed lower predictive values (S7E Fig).
Publication year.
No significant difference was found between studies published in 2016−2020 (OR = 0.77, 95% CI: 0.22–2.74, I2 = 89.7%) and those published in 2021−2024 (OR = 0.81, 95% CI: 0.78–0.84, I2 = 61.6%) (χ²₁ = 0.20, p = 0.656) (S7F Fig).
Publication bias and sensitivity analysis.
Visual inspection of the funnel plot showed symmetrical distribution of studies (S7G Fig), and Egger’s test confirmed no significant publication bias (t = 0.54, df = 11, p = 0.5967; bias estimate = 1.0742, SE = 1.7913). Sensitivity analysis using the leave-one-out method demonstrated robust results, with the pooled OR ranging from 0.79 (95% CI: 0.76–0.82) to 0.80 (95% CI: 0.77–0.84) when individual studies were sequentially removed, indicating that no single study disproportionately influenced the overall results (S7H Fig).
Calibration was assessed in nine studies: six reported good calibration (Hosmer–Lemeshow p > 0.05), and three used calibration curves, two of which showed moderate deviation (slope < 1.0, intercept > 0). One high-discrimination model exhibited poor calibration, suggesting overfitting.
Model validation
Fifteen studies underwent internal validation (random split: n = 7; cross-validation: n = 4; bootstrapping: n = 2), and 11 included both internal and external validation.
Risk of bias and applicability
The “Participants” domain was assessed as having a high risk of bias in all studies, primarily because retrospective designs may introduce information bias due to the unsystematic collection of predictor and outcome data, which is not ideal for prognostic modeling. Prospective cohort studies, which follow a longitudinal temporal relationship between predictors and outcomes, are considered the optimal study design, as they capture disease progression in its natural state.
For the “Predictors” domain, two studies were assessed as having a low risk of bias, particularly those utilizing prospective cohort designs, where predictors were measured before outcomes occurred. However, 22 studies were assessed as having an unclear risk of bias because they did not specify whether predictors were evaluated independently of outcome knowledge.
In the “Outcome” domain, two studies were assessed as having a low risk of bias due to the retrospective design, where outcomes were determined before predictor measurements, potentially linking outcome determination to the predictor. The remaining 22 studies were assessed as having an unclear risk of bias due to insufficient information on whether predictor data was available at the time of outcome determination.
In the “Analysis” domain, 22 studies were assessed as having a high risk of bias. Two studies had insufficient sample sizes, failing to meet the criterion that the number of clinical outcome events should be 20 times greater than the number of potential predictors. Fifteen studies converted continuous variables into categorical ones, leading to potential information loss, abrupt changes in predictions near thresholds, reduced statistical efficacy, and decreased result credibility. Three studies directly deleted missing data, which may have introduced selection bias and resulted in information loss. Additionally, 21 studies filtered predictors based on univariate analysis, potentially overemphasizing statistical significance while neglecting non-significant variables that might still contribute to prediction.
Regarding applicability, in the “Patients” domain, two studies included patients with choledocholithiasis undergoing ERCP, one study included patients with malignant biliary obstruction undergoing ERCP, and one study included patients with early onset hyperamylasemia after ERCP treatment. These studies had a poor risk of applicability to the broader study population. The remaining 20 studies were assessed as having a low risk of applicability. The “Predictors” and “Outcome” domains were assessed as having low concerns about applicability in all studies (Table 4).
While PROBAST remains our primary tool given its specificity for prediction models, supplementary assessment with NOS revealed that all 24 studies met high-quality criteria (NOS ≥ 7). The study-specific NOS scores are available in S4 Table.
Certainty of evidence (GRADE)
GRADE assessment indicated an overall moderate certainty of evidence. Evidence profiles generated by GRADEpro software are available in S5 Table.
Risk of bias.
The included prediction models adhered to rigorous standards in predictor measurement and outcome ascertainment. However, some studies failed to report model calibration details, leading to an overall moderate risk of bias (downgraded by one level).
Indirectness.
While most models were derived from tertiary center cohorts, limiting generalizability to primary care settings, the universal applicability of predictor variables (e.g., procedural difficulty scores) justified no downgrade.
Inconsistency.
Substantial heterogeneity in model discrimination was observed (I2 = 82%), likely due to variations in endoscopist expertise (downgraded by one level).
Discussion
Principal findings
This systematic review and meta-analysis synthesized evidence from 24 studies on post-ERCP pancreatitis (PEP) prediction models. Three key findings emerged: (i) the pooled incidence of PEP highlights its substantial clinical burden, particularly among high-risk patient groups; (ii) several predictors—including pancreatic duct cannulation, difficult cannulation, pancreatography, female sex, pancreatic duct guidewire use, history of pancreatitis, total bilirubin, albumin, and choledocholithiasis—show consistent and significant associations with PEP risk; (iii) while many models achieved good discrimination (AUC > 0.80), calibration reporting and external validation were often insufficient, and overall methodological quality varied considerably.
PEP incidence and clinical implications
Our pooled estimates confirm that PEP remains a frequent complication of ERCP, with higher incidence in high-risk groups. This reinforces the need for risk-stratified prevention strategies and justifies the use of predictive models to identify patients who may benefit most from prophylactic interventions. Although incidence estimates varied, much of the heterogeneity can be explained by procedural factors, operator experience, and patient selection criteria—highlighting the importance of adjusting models for these variables.
Predictors and underlying mechanisms
Across the included studies, the most frequent predictors were pancreatic duct cannulation, difficult cannulation, pancreatography, female sex, history of pancreatitis, elevated total bilirubin, low albumin, pancreatic duct guidewire use, longer procedure time, and choledocholithiasis.
Mechanistically, these factors may cause mechanical or chemical injury to the pancreatic duct, induce papillary congestion or Oddi sphincter spasm, increase intraductal pressure, or promote premature activation of pancreatic enzymes. For example, repeated or difficult cannulation can exacerbate mucosal edema and obstruct pancreatic fluid outflow, while contrast injection during pancreatography may create high intraductal pressure, leading to reflux of pancreatic juice into the parenchyma and triggering autodigestion [39,40]. Female patients may be more susceptible due to a higher prevalence of sphincter of Oddi dysfunction [41]. A history of pancreatitis may reflect latent injury to the pancreaticobiliary system, increasing vulnerability to procedure-related trauma. Low albumin levels may indicate impaired protein metabolism and reduced tissue resilience, whereas elevated total bilirubin may reflect pancreaticobiliary dysfunction; both have been variably linked to PEP risk [42–44], though inconsistent findings suggest the need for further validation in large, multicenter studies.
Notably, our analysis and prior systematic reviews consistently identify pancreatic duct cannulation and female sex as robust predictors [45,46], with the latter possibly linked to sex-specific differences in sphincter of Oddi tone and reactivity [47–49]. Beyond individual predictors, our findings highlight the importance of synergistic effects: combined mechanical insults from cannulation, pancreatography, and guidewire use may produce cumulative injury, substantially elevating PEP risk compared to any single factor alone [50–52]. Prolonged procedure time may further amplify these effects, especially in patients with compromised metabolic or structural resilience [53].
Given these observations, future predictive models should explicitly incorporate such interactions, either through interaction terms in traditional regression or via non-linear modeling approaches, to better capture the complex interplay of patient-, procedure-, and operator-related factors and improve risk stratification accuracy [54,55].
Predictive model performance and methodological quality
Across the included studies, model discrimination exhibited wide variability (AUC range: 0.560–0.915), with 58.3% (14/24) achieving an AUC greater than 0.80. Higher-performing models generally incorporated multi-domain predictors—encompassing biochemical, procedural, and anatomical variables—whereas lower-performing models relied predominantly on demographic characteristics and basic laboratory parameters. Nevertheless, discrimination alone does not ensure accurate absolute risk estimation. Calibration, which evaluates the concordance between predicted and observed probabilities, was reported in only nine studies, and calibration plots—currently the most informative method—were seldom employed. Several models with high AUC values demonstrated poor calibration, indicating potential overfitting and limited clinical transportability.
Our PROBAST assessment identified several recurring methodological limitations: (i) Many models failed to satisfy the widely recommended threshold of at least 20 events per variable (EPV), which is critical to reduce overfitting and enhance generalizability. For instance, a model with 10 predictors should ideally be based on ≥200 outcome events to ensure statistical stability; (ii) Continuous predictors were frequently dichotomized or converted into categorical variables, leading to loss of information and reduced predictive power. More robust strategies include retaining continuous variables in their original form, applying restricted cubic splines, or determining optimal cutoffs using statistical indices such as the Youden index from ROC analysis; (iii) Complete-case analysis was frequently adopted, which risks introducing bias and diminishing statistical efficiency. Multiple imputation, accompanied by sensitivity analyses, is preferable to maximize data utilization and maintain the validity of results; (iv) The common practice of univariate pre-screening can exclude clinically relevant variables, as statistical insignificance in univariate analysis does not preclude predictive value in multivariable contexts. Candidate predictors should be selected based on prior literature, established pathophysiological mechanisms, and clinical reasoning, alongside statistical considerations; (v) Sixteen studies did not perform internal validation; bootstrap resampling is recommended, particularly for small datasets, as it avoids the sample reduction inherent to cross-validation. External validation was conducted in only 12 studies, limiting generalizability to broader populations and diverse clinical settings.
When integrated with our meta-analytic findings, the pooled estimates of model performance, the overall incidence of PEP, and the effect sizes of individual predictors collectively reveal marked heterogeneity in model design and quality. These observations highlight the necessity for standardized methodological frameworks, rigorous internal and external validation, and transparent reporting to facilitate the development of robust, clinically applicable prediction models for PEP.
Methodological implications
These findings directly support the appropriateness of our systematic review and meta-analysis approach. Although predictor heterogeneity is inevitable, pooling allowed us to (i) identify stable, high-impact predictors; (ii) quantify model performance; and (iii) estimate PEP incidence—three complementary outputs that collectively inform both clinical application and model optimization.
Clinical vpractice recommendations
Model development.
Incorporate diverse predictors, account for potential interactions, and retain continuous variables where possible. We recommend building models using multicenter datasets that encompass a wide range of potential risk factors, to enhance representativeness and robustness.
Advanced methods.
Integrate AI/ML techniques (e.g., random forest, support vector machines) alongside traditional logistic regression to capture complex relationships. Future research should explore whether combining traditional biostatistical approaches with AI/ML can yield better predictive performance than either method alone.
Validation.
Conduct temporal, geographic, and domain-specific external validations using independent datasets. Among the included studies, 12 performed external validation, highlighting its feasibility.
Implementation.
Present models via user-friendly tools (web calculators, mobile apps) and integrate with electronic health record systems to automate data entry and support real-time decision-making. Notably, one study transformed its model into an interactive online visualization platform for clinical use, which could serve as a practical example for promoting clinical adoption.
Limitation
Our analysis is limited by the predominance of observational studies, variable reporting quality, and frequent absence of calibration or external validation in included models. Restricting to English and Chinese literature may have excluded relevant studies. Furthermore, high overall risk of bias in PROBAST assessments means that most models require further refinement and validation before routine clinical use.
Conclusion
This study included 24 studies, encompassing 26 PEP risk prediction models. The results showed that most of the existing prediction models had good predictive performance, but the overall risk of bias was high. It is recommended that future research strictly follow the TROPID guidelines, conduct multicenter and large-sample studies to develop high-quality predictive model research to provide reference for clinical decision-making.
Supporting information
S4 Table. Risk of bias assessment using the Newcastle Ottawa Scale.
https://doi.org/10.1371/journal.pone.0332378.s004
(DOCX)
S5 Table. GRADE assessment for post-endoscopic retrograde cholangiopancreatography pancreatitis.
https://doi.org/10.1371/journal.pone.0332378.s005
(DOCX)
S1 Fig. Forest plot of Post-ERCP Pancreatitis incidence: a meta-analysis.
https://doi.org/10.1371/journal.pone.0332378.s006
(DOCX)
S2 Fig. Forest plot of Post-ERCP Pancreatitis incidence: subgroup meta-analysis by study design.
https://doi.org/10.1371/journal.pone.0332378.s007
(DOCX)
S3 Fig. Forest plot of Post-ERCP Pancreatitis incidence: subgroup meta-analysis by diagnostic criteria.
https://doi.org/10.1371/journal.pone.0332378.s008
(DOCX)
S4 Fig. Forest plot of Post-ERCP Pancreatitis incidence: subgroup meta-analysis by geographic region.
https://doi.org/10.1371/journal.pone.0332378.s009
(DOCX)
S5 Fig. Forest plot of Post-ERCP Pancreatitis incidence: subgroup meta-analysis by validation method.
https://doi.org/10.1371/journal.pone.0332378.s010
(DOCX)
S6 Fig. Forest plot of the meta-analysis on risk factors for Post-ERCP Pancreatitis.
https://doi.org/10.1371/journal.pone.0332378.s011
(DOCX)
S7 Fig. Forest plot of the meta-analysis on predictive performance for Post-ERCP Pancreatitis.
https://doi.org/10.1371/journal.pone.0332378.s012
(DOCX)
References
- 1. Dumonceau J-M, Delhaye M, Tringali A, Arvanitakis M, Sanchez-Yague A, Vaysse T, et al. Endoscopic treatment of chronic pancreatitis: European Society of Gastrointestinal Endoscopy (ESGE) Guideline - Updated August 2018. Endoscopy. 2019;51(2):179–93. pmid:30654394
- 2. ASGE Standards of Practice Committee, Buxbaum JL, Abbas Fehmi SM, Sultan S, Fishman DS, Qumseya BJ, et al. ASGE guideline on the role of endoscopy in the evaluation and management of choledocholithiasis. Gastrointest Endosc. 2019;89(6):1075-1105.e15. pmid:30979521
- 3. Cianci P, Restini E. Management of cholelithiasis with choledocholithiasis: endoscopic and surgical approaches. World J Gastroenterol. 2021;27(28):4536–54. pmid:34366622
- 4. Dumonceau J-M, Kapral C, Aabakken L, Papanikolaou IS, Tringali A, Vanbiervliet G, et al. ERCP-related adverse events: European Society of Gastrointestinal Endoscopy (ESGE) Guideline. Endoscopy. 2020;52(2):127–49. pmid:31863440
- 5. ASGE Standards of Practice Committee, Chandrasekhara V, Khashab MA, Muthusamy VR, Acosta RD, Agrawal D, et al. Adverse events associated with ERCP. Gastrointest Endosc. 2017;85(1):32–47. pmid:27546389
- 6. Sajid MS, Khawaja AH, Sayegh M, Singh KK, Philipose Z. Systematic review and meta-analysis on the prophylactic role of non-steroidal anti-inflammatory drugs to prevent post-endoscopic retrograde cholangiopancreatography pancreatitis. World J Gastrointest Endosc. 2015;7(19):1341–9. pmid:26722616
- 7. Zhang QQ, Wang GZ, Lu QF. Relationship between NLR and PLR on acute severe pancreatitis after ERCP surgery. J Mol Diagnostics Therapy. 2022;14(11):1918–21.
- 8. Moons KGM, Wolff RF, Riley RD, Whiting PF, Westwood M, Collins GS, et al. PROBAST: a tool to assess risk of bias and applicability of prediction model studies: explanation and elaboration. Ann Intern Med. 2019;170(1):W1–33. pmid:30596876
- 9. Stang A. Critical evaluation of the Newcastle-Ottawa scale for the assessment of the quality of nonrandomized studies in meta-analyses. Eur J Epidemiol. 2010;25(9):603–5. pmid:20652370
- 10. Schünemann HJ, Oxman AD, Brozek J, Glasziou P, Jaeschke R, Vist GE, et al. Grading quality of evidence and strength of recommendations for diagnostic tests and strategies. BMJ. 2008;336(7653):1106–10. pmid:18483053
- 11. Friedland S, Soetikno RM, Vandervoort J, Montes H, Tham T, Carr-Locke DL. Bedside scoring system to predict the risk of developing pancreatitis following ERCP. Endoscopy. 2002;34(6):483–8. pmid:12048633
- 12. DiMagno MJ, Spaete JP, Ballard DD, Wamsteker E-J, Saini SD. Risk models for post-endoscopic retrograde cholangiopancreatography pancreatitis (PEP): smoking and chronic liver disease are predictors of protection against PEP. Pancreas. 2013;42(6):996–1003. pmid:23532001
- 13.
Fang JF. The analysis of the risk factors of post-ERCP pancreatitis and the establishment of predictive models. [Master’s thesis]. Zhejiang University; 2016.
- 14. Wan XL. Analysis of the risk factors for post-ERCP pancreatitis and assessment its predict effect in patients of gastroenterology department. Chinese J Integrat Traditional Western Med Digestion. 2018;26(08):677–80.
- 15. Chiba M, Kato M, Kinoshita Y, Shimamoto N, Tomita Y, Abe T, et al. The milestone for preventing post-ERCP pancreatitis using novel simplified predictive scoring system: a propensity score analysis. Surg Endosc. 2021;35(12):6696–707. pmid:33258029
- 16.
Dou L. Develop and validate a clinical model for predicting the risk of post-ERCP pancreatitis. [Master’s thesis]. Chongqing Medical University; 2021.
- 17.
Wang C. Analysis the risk factors of post-ERCP pancreatitis and establishment of risk scoring system. [Master’s thesis]. Dalian Medical University; 2021.
- 18.
Zhang M. Nomogram to predict the risk of pancreatitis after first ERCP. [Master’s thesis]. Nanchang University; 2022.
- 19.
Zheng RH. Development and validation of a risk prediction model and scoring system for post-endoscopic retrograde cholangiopancreatography pancreatitis. [Master’ s thesis]. Nanjing Medical University; 2021.
- 20. Park CH, Park SW, Yang MJ, Moon SH, Park DH. Pre- and post-procedure risk prediction models for post-endoscopic retrograde cholangiopancreatography pancreatitis. Surg Endosc. 2022;36(3):2052–61. pmid:34231067
- 21. Fujita K, Yazumi S, Uza N, Kurita A, Asada M, Kodama Y, et al. New practical scoring system to predict post-endoscopic retrograde cholangiopancreatography pancreatitis: development and validation. JGH Open. 2021;5(9):1078–84. pmid:34584978
- 22. Zhang X, Zhang JD, Pei ZJ. Effect of nomogram model of gallbladder state on pancreatitis after ERCP. J Lanzhou Univ (Medical Sciences). 2022;48(01):44–9.
- 23. Fu Z, Song J, Pi Y, Sun X, Liu M, Xiao Z, et al. A risk prediction model for post-endoscopic retrograde cholangiopancreatography pancreatitis after stent insertion for malignant biliary obstruction: development and validation. Dig Dis Sci. 2023;68(4):1574–84. pmid:35989385
- 24. Huang YC. Development and validation of a risk prediction model for post-ERCP pancreatitis. [Master’s thesis]. Henan University; 2022.
- 25. Archibugi L, Ciarfaglia G, Cárdenas-Jaén K, Poropat G, Korpela T, Maisonneuve P, et al. Machine learning for the prediction of post-ERCP pancreatitis risk: A proof-of-concept study. Dig Liver Dis. 2023;55(3):387–93. pmid:36344369
- 26. Ma YY, Ding ZY, Lu XP. Construction and validation of a risk predictive model for post endoscopic retrograde cholangiopancreatography pancreatitis. China J Endoscopy. 2023;29(12):65–71.
- 27. Yao JH, Chai B, Cao XJ. Risk factors of complication of pancreatitis after endoscopic retrograde cholangiopancreatography and construction of prediction model. Chinese Nursing Res. 2023;37(5):814–8.
- 28. Qin YJ, Sha J, Zhu MH. Construction and validation of a visual nomograph model for pancreatitis after endoscopic retrograde cholangiopancreatography. J Qiqihar Med Univ. 2023;44(18):1746–50.
- 29. Wang SJ, Chen J, Zhang GF. Construction of risk factor regression equation for postoperative pancreatitis after ERCP operation. J Hepatopancreatobiliary Surg. 2023;35(10):595–601.
- 30.
Chen TJ. Analysis of factors and predictive model for the development of pancreatitis in patients with early hyperamylasemia after ERCP. [Master’s thesis]. Bengbu Medical College; 2023.
- 31. Takahashi H, Ohno E, Furukawa T, Yamao K, Ishikawa T, Mizutani Y, et al. Artificial intelligence in a prediction model for post-endoscopic retrograde cholangiopancreatography pancreatitis. Dig Endosc. 2024;36(4):463–72. pmid:37448120
- 32. Fukuda R, Hakuta R, Nakai Y, Hamada T, Takaoka S, Tokito Y, et al. Development and external validation of a nomogram for prediction of post-endoscopic retrograde cholangiopancreatography pancreatitis. Pancreatology. 2023;23(7):789–96. pmid:37666733
- 33.
Zhang Y. Identification of risk factors for post-ERCP pancreatitis and validation and establishment of prediction models. [Master’s thesis]. Xian Medical College; 2023.
- 34. Yan C, Zheng J, Tang H, Fang C, Zhu J, Feng H, et al. Prediction for post-ERCP pancreatitis in non-elderly patients with common bile duct stones: a cross-sectional study at a major Chinese tertiary hospital (2015-2023). BMC Med Inform Decis Mak. 2024;24(1):143. pmid:38807169
- 35. Cotton PB, Lehman G, Vennes J, Geenen JE, Russell RC, Meyers WC, et al. Endoscopic sphincterotomy complications and their management: an attempt at consensus. Gastrointest Endosc. 1991;37(3):383–93. pmid:2070995
- 36. Banks PA, Bollen TL, Dervenis C, Gooszen HG, Johnson CD, Sarr MG, et al. Classification of acute pancreatitis--2012: revision of the Atlanta classification and definitions by international consensus. Gut. 2013;62(1):102–11. pmid:23100216
- 37. Du YQ, Li WQ, Mao EQ. Chinese consensus on the multidisciplinary treatment (MDT) of acute pancreatitis. J Clin Hepatol. 2015;31(11):1770–5.
- 38. Cotton PB, Eisen GM, Aabakken L, Baron TH, Hutter MM, Jacobson BC, et al. A lexicon for endoscopic adverse events: report of an ASGE workshop. Gastrointest Endosc. 2010;71(3):446–54. pmid:20189503
- 39. Nakai Y, Kusumoto K, Itokawa Y, Inatomi O, Bamba S, Doi T, et al. Emergency endoscopic retrograde cholangiopancreatography did not increase the incidence of postprocedural pancreatitis compared with elective cases: a prospective multicenter observational study. Pancreas. 2022;51(1):41–7. pmid:35195594
- 40. Wang J, Su J, Lu Y, Zhou H, Gong B. A randomized control study to investigate the application of Ulinastatin-containing contrast medium to prevent post-ERCP pancreatitis. Hepatogastroenterology. 2014;61(136):2391–4. pmid:25699389
- 41. Freeman ML, DiSario JA, Nelson DB, Fennerty MB, Lee JG, Bjorkman DJ, et al. Risk factors for post-ERCP pancreatitis: a prospective, multicenter study. Gastrointest Endosc. 2001;54(4):425–34. pmid:11577302
- 42. Morales SJ, Sampath K, Gardner TB. A review of prevention of post-ERCP pancreatitis. Gastroenterol Hepatol (N Y). 2018;14(5):286–92.
- 43. Park C-H, Paik WH, Park ET, Shim CS, Lee TY, Kang C, et al. Aggressive intravenous hydration with lactated Ringer’s solution for prevention of post-ERCP pancreatitis: a prospective randomized multicenter clinical trial. Endoscopy. 2018;50(4):378–85. pmid:29237204
- 44. Zhang Z-F, Duan Z-J, Wang L-X, Zhao G, Deng W-G. Aggressive hydration with lactated ringer solution in prevention of postendoscopic retrograde cholangiopancreatography pancreatitis: a meta-analysis of randomized controlled trials. J Clin Gastroenterol. 2017;51(3):e17–26. pmid:28178088
- 45. Beran A. Predictors of post-endoscopic retrograde cholangiopancreatography pancreatitis: a comprehensive systematic review and meta-analysis. Clin Gastroenterol Hepatol.
- 46. Sperna Weiland CJ, Akshintala VS, Singh A, Buxbaum J, Choi J-H, Elmunzer BJ, et al. Preventive measures and risk factors for Post-ERCP pancreatitis: a systematic review and individual patient data meta-analysis. Dig Dis Sci. 2024;69(12):4476–88. pmid:39500841
- 47. Masci E, Mangiavillano B, Luigiano C, Bizzotto A, Limido E, Cantù P, et al. Comparison between loop-tip guidewire-assisted and conventional endoscopic cannulation in high risk patients. Endosc Int Open. 2015;3(5):E464-70. pmid:26528503
- 48.
Park SM. Sex/gender differences in pancreatic and biliary diseases. In: Kim N, Kim N, eds. Sex/gender-specific medicine in clinical areas. Singapore: Springer Nature Singapore; 2024: 219–30.
- 49. Tse F, Yuan Y, Bukhari M, Leontiadis GI, Moayyedi P, Barkun A. Pancreatic duct guidewire placement for biliary cannulation for the prevention of post-endoscopic retrograde cholangiopancreatography (ERCP) pancreatitis. Cochrane Database Syst Rev. 2016;2016(5):CD010571. pmid:27182692
- 50. Chiriac S, Sfarti CV, Stanciu C, Cojocariu C, Zenovia S, Nastasa R, et al. The relation between post-endoscopic retrograde cholangiopancreatography pancreatitis and different cannulation techniques: the experience of a high-volume center from North-Eastern Romania. Life (Basel). 2023;13(6):1410. pmid:37374192
- 51. Fung BM, Pitea TC, Tabibian JH. Difficult biliary cannulation in endoscopic retrograde cholangiopancreatography: definitions, risk factors, and implications. Eur Med J Hepatol. 2021;9(1):64–72. pmid:34621527
- 52. Han S, Zhang J, Durkalski-Mauldin V, Foster LD, Serrano J, Coté GA, et al. Impact of difficult biliary cannulation on post-ERCP pancreatitis: secondary analysis of the stent versus indomethacin trial dataset. Gastrointest Endosc. 2025;101(3):617–28. pmid:39389431
- 53. Tang P, Zhang J, Yang J, Guo S, Wang H. Risk factors analysis of short-term effect after pancreatoduodenectomy. Chongqing Med J. 2023;52(21):3232–8.
- 54. Archibugi L, Ciarfaglia G, Cárdenas-Jaén K, Poropat G, Korpela T, Maisonneuve P, et al. Machine learning for the prediction of post-ERCP pancreatitis risk: a proof-of-concept study. Digestive Liver Dis. 2022;55.
- 55. Ding X, Zhang F, Wang Y. Risk factors for post-ERCP pancreatitis: a systematic review and meta-analysis. Surgeon. 2015;13(4):218–29. pmid:25547802