Toward Reproducible and Transportable Anastomotic Leak Intelligence: Benchmark Design, Distribution Shift, and Decision Reliability

Abstract

Postoperative complications remain one of the most analytically demanding targets in clinical prediction because they emerge from interacting physiologic, procedural, and organizational processes. Anastomotic leak is a particularly difficult example. It is uncommon enough to create statistical imbalance, serious enough to motivate intervention, and clinically heterogeneous enough that retrospective labels do not always map cleanly onto prospective decisions. For these reasons, model development alone does not resolve the central translational problem. What is needed is a principled account of how predictive systems should be specified, benchmarked, evaluated, and maintained when the intended use environment is a hospital rather than a static test set. This paper develops a technical framework for the transition from isolated proof-of-concept performance to reproducible and transportable surgical prediction infrastructure. The analysis treats leak prediction as a layered problem involving phenotype construction, observation process modeling, institution-sensitive generalization, rare-event calibration, utility-aware operating rules, and prospective evaluation under behavioral feedback. Rather than asking only whether a classifier separates cases from controls, the paper asks which statistical objects remain stable across centers, which measures of performance retain operational meaning under prevalence variation, and which validation designs can distinguish genuine predictive robustness from local data familiarity. The proposed framework argues for benchmark designs that preserve semantic traceability of features, explicit separation of discrimination from calibration, hierarchical treatment of site heterogeneity, and deployment governance that treats drift as expected rather than exceptional. The aim is neither to reject machine learning nor to assume its inevitability, but to specify the conditions under which surgical prediction could become a durable component of perioperative intelligence.