Abstract
Anastomotic leak remains one of the most consequential complications after gastrointestinal surgery, yet its risk is seldom uniform across procedures, hospitals, and perioperative care environments. Aggregate estimates are clinically convenient but analytically coarse because they compress biologically distinct operations and institutionally distinct treatment settings into a single average. That averaging can obscure where risk is intrinsically high, where it is context-dependent, and where observed variation is largely a by-product of case mix. This study examined heterogeneity of leak risk across major gastrointestinal surgery classes and hospital care settings using a retrospective multicenter electronic health record cohort. Adults undergoing right colectomy, left or sigmoid colectomy, low anterior resection, subtotal or total colectomy, small-bowel resection with anastomosis, gastrectomy, or esophagectomy were included. Hospital care setting was classified as tertiary referral, regional teaching, community, or urban safety-net, with immediate postoperative disposition modeled separately. We estimated patient-level leak probability within 30 days using a hierarchical logistic framework with nonlinear effects, surgery-specific random terms, setting-specific random terms, and cross-level interaction structure. The cohort comprised 48,917 index procedures with 3,218 leaks, yielding an overall event rate of 6.6\%. Crude leak incidence varied from 3.5\% after right colectomy to 13.2\% after esophagectomy. After risk standardization, surgery class remained the dominant source of heterogeneity, but setting-specific differences persisted for technically demanding distal colorectal and foregut procedures. Low anterior resection and esophagectomy showed the largest setting-dependent gradients, whereas right colectomy exhibited limited contextual dispersion. A transportable hierarchical model achieved stable discrimination and calibration across held-out hospitals. The findings suggest that clinically useful leak surveillance should not rely on a single pooled threshold. Procedure-sensitive and setting-aware risk estimation appears more appropriate for postoperative escalation, diagnostic timing, and resource allocation.