Abstract
We conducted a multi-center empirical study to examine whether external validation, rather than source-domain fit, determines the main limit on translational credibility for early prediction of colorectal anastomotic leak. Adult patients undergoing colorectal resection with primary anastomosis were identified in twelve hospitals from 2017 through 2025. Landmark prediction was performed at the end of postoperative day 1 using preoperative, intraoperative, and early postoperative structured variables. Three model families were compared: penalized logistic regression, gradient boosting, and a hospital-adaptive generalized additive model with partial pooling. Performance was evaluated using temporal validation, geographic validation, calibration, decision-curve analysis, and threshold stability across institutions. The study cohort included 21,408 operations, of which 1,472 were followed by clinically significant leak within postoperative day 10. The highest internal discrimination was achieved by the source-optimized boosting model, but that model showed marked external calibration distortion and large site-to-site threshold instability. The hospital-adaptive model had slightly lower internal discrimination yet materially better external calibration, lower threshold regret, and more stable net benefit. Observation-intensity variables improved source-domain performance but accounted for a substantial fraction of transport failure. These findings indicate that the principal translational obstacle was not absence of predictive signal. It was instability of score meaning once the model left the environments that generated the original clinical records.