Risk prediction models for delayed gastric emptying in patients after pancreaticoduodenectomy: a systematic review and meta-analysis.

Objectives: Although several risk prediction models for delayed gastric emptying (DGE) after pancreaticoduodenectomy (PD) have been developed, their suitability for clinical application and future studies remains uncertain. This study aimed to systematically review published studies on risk prediction models for DGE after PD.

Design: A systematic review and meta-analysis.

Data sources: Chinese databases (such as China National Knowledge Infrastructure, VIP, SinoMed and Wanfang), PubMed, Web of Science, Cochrane Library, Embase and Scopus were searched from their inception to May 2025 to identify studies describing DGE prediction models.

Eligibility criteria: All observational studies that developed DGE prediction models for post-pancreatectomy patients were included. Eligible models were required to incorporate at least two predictive variables. Studies were excluded if they were unpublished, not available in English or Chinese or lacked sufficient methodological details regarding study design, model development or statistical analysis.

Data extraction and synthesis: Two reviewers independently screened eligible studies and extracted relevant data. The risk of bias was assessed using the Prediction Model Risk of Bias Assessment Tool. A random-effects meta-analysis was conducted to estimate the pooled area under the curve (AUC). Heterogeneity among studies was evaluated using the I² statistic and Cochran's Q test, and publication bias was assessed through Egger's regression test and funnel plot symmetry. Additionally, to evaluate the robustness of the findings, leave-one-out sensitivity analyses were conducted.

Results: This systematic review included 12 studies (n=24 453), with reported incidences of DGE after PD ranging from 11.7% to 37.9%. Most studies had a high overall bias risk, primarily due to retrospective designs and inadequate model validation. Pooled analysis showed moderate predictive accuracy (AUC=0.728, 95% CI 0.678 to 0.778) with moderate heterogeneity (I² = 32.6%). No significant publication bias was detected (Egger's p=0.12), and sensitivity analyses supported the robustness of the findings (AUC=0.70-0.77).

Conclusions: Current risk prediction models for DGE after PD remain in the preliminary stages of development and generally exhibit a high risk of bias. Future efforts should focus on developing models with strong prediction performance, low bias risk and convenience for clinical application.