Competing Risk Trajectory Modeling of Anastomotic Leak Emergence Across Gastrointestinal Procedure Families and Postoperative Care Environments

Abstract

Postoperative anastomotic leak is usually summarized as a binary outcome, but its clinical significance depends not only on whether it occurs, but also on when it becomes detectable and under what surveillance conditions it is recognized. A leak identified on postoperative day 3 in a highly monitored environment is not operationally equivalent to a leak first detected after several days on a lower-observation ward. This study examined temporal heterogeneity in leak emergence across gastrointestinal procedure families and postoperative care environments using a retrospective multicenter electronic health record cohort. Adult patients undergoing gastrointestinal operations with restoration of continuity were followed through daily postoperative records for 21 days and outcome ascertainment for 30 days. Procedures were grouped into proximal colonic, distal colonic, low pelvic colorectal, small-bowel, and upper gastrointestinal reconstruction families. Care environment was defined by the joint combination of first postoperative bed pathway and hospital rescue capability. The primary endpoint was clinically consequential anastomotic leak, with discharge and death before leak treated as competing events. The analytic cohort included 41,286 procedures, of which 2,521 developed leak within 30 days, corresponding to an overall incidence of 6.1\%. Low pelvic colorectal and upper gastrointestinal reconstructions had the highest crude event rates and the latest median day of detection. After covariate standardization, proximal colonic and small-bowel procedures showed limited temporal variation across care environments, whereas low pelvic colorectal and upper gastrointestinal procedures displayed substantial shifts in both cumulative incidence and detection timing. Ward-first pathways in standard rescue environments were associated with later recognition and higher post-detection rescue intensity for these procedures. A dynamic competing-risk model with low-rank procedure-by-environment interactions outperformed a static postoperative baseline model in discrimination, calibration, and cross-environment transport. The findings suggest that leak surveillance should be calibrated to procedural trajectory and monitoring context, rather than treated as a single undifferentiated binary risk problem.