External Validation and Cross-Hospital Generalizability of Postoperative Anastomotic Leak Prediction Models

Abstract

Anastomotic leak remains one of the most consequential complications after colorectal resection because it alters recovery trajectories, prolongs hospitalization, increases reintervention, and complicates decisions about diversion, surveillance, and adjuvant treatment timing. Risk models have therefore attracted sustained attention, yet many published systems are developed in single institutions and are judged primarily by internal discrimination. That pattern leaves unanswered whether the same model preserves accuracy when moved across hospitals with different case mix, measurement routines, perioperative pathways, and operative styles. The present study examined generalizability and external validation of postoperative leak prediction in East Asia using a retrospective multi-center cohort derived from fifteen hospitals across mainland China, Japan, South Korea, Taiwan, and Hong Kong. Adult patients undergoing colorectal resection with primary anastomosis were included, and leak within 30 days was ascertained using a chart-validated composite outcome. Three modeling strategies were compared: pooled elastic-net logistic regression, gradient boosting, and a hierarchical transport model with hospital-specific random effects and post-hoc calibration updating. The final analytic sample contained 32{,}964 patients, with 2{,}196 leaks for an overall incidence of 6.7\%. Internal temporal performance was strongest for gradient boosting, but the hierarchical transport model showed the most stable geographic performance. In previously unseen hospitals, the hierarchical approach achieved an AUC of 0.781, PR-AUC of 0.212, and Brier score of 0.055, improving further after lightweight local recalibration. Performance decline under country holdout was modest but nonuniform, with the largest degradation observed in low-volume rectal surgery settings. The results indicate that transport-aware modeling and site-sensitive recalibration matter more for external readiness than maximal internal discrimination alone.}