Artificial Intelligence Benchmarking of Sepsis Onset Predictors Under Unmeasured Ward-Transfer Censoring in Electronic Health Records: A Healthcare Safety-Robustness Evaluation

Sepsis onset prediction models are benchmarked on electronic health record (EHR) cohorts in which a substantial fraction of patients are silently transferred between wards before the prediction horizon closes. Because ward transfer is a competing event that is neither recorded as a censoring time nor modelled in standard benchmarking pipelines, reported discrimination and calibration statistics are computed on a survivor-biased risk set. the present analysis formalises the resulting bias as an unmeasured competing-risk censoring problem, derives the asymptotic bias of the area under the receiver operating characteristic curve (AUROC) and of the calibration slope under a declared variance structure, and proposes a safety-robustness benchmarking protocol that reports a bounded worst-case degradation envelope rather than a single point estimate. The analysis is a design and power study, not an observational one: no patient data are used or claimed. Operating characteristics are characterised across a declared grid of 8, 16, 24 and 32 clusters, effect sizes of 3, 5 and 8 percentage points, an intraclass correlation of 0.64, a between-cluster standard deviation of 6.0 and a residual standard deviation of 4.5, with 2000 replications per design cell. Under these parameters, 52 clusters are required to detect a 5 percentage point effect at 80% power. The contribution is a formal account of why ward-transfer censoring inflates apparent sepsis-predictor performance and a benchmarking protocol that makes the inflation visible and bounded.