Abstract
Postoperative risk models are commonly summarized by a single discrimination score, yet that convention hides the practical structure of error. In anastomotic leak prediction, this omission is consequential because the event is infrequent, the clinical stakes are high, and modest changes in precision or false-alarm burden can alter whether a model is operationally credible. This paper presents an empirical secondary analysis of an early postoperative leak prediction experiment based on structured perioperative records. Rather than asking only which classifier attains the best overall ranking, the study examines where informative signal is concentrated, how false positives accumulate across model families, and how threshold behavior reshapes the balance between detection and workload. The analysis treats reported test-set performance as an observable object from which implied alert rates, event enrichment, and miss-to-capture trade-offs can be derived. Particular attention is given to the relation between nonlinear modeling, probability quality, and the asymmetry created by low event prevalence. The study finds that model differences are not exhausted by changes in area under the receiver operating characteristic curve. Some models compress a meaningful share of leak events into a relatively small alert surface, whereas others recover additional cases only by expanding false-positive volume substantially. The dominant predictors are also notable: early inflammatory and physiologic response variables outweigh many static baseline factors, suggesting that the first postoperative day already contains a concentrated but imperfect signature of later leak. The findings support a more granular empirical language for surgical prediction, one centered on error topology and operational yield rather than on discrimination alone.