Artificial Intelligence Benchmarking of Sepsis Onset Prediction from Nurse-Collected Vital Signs Under Unmeasured Shift Handover Transcription Lag: A Healthcare Early-Warning Evaluation

Sepsis early-warning models trained on nurse-collected vital signs inherit a temporal defect that benchmarking practice has not yet named: the interval between a bedside observation and its transcription into the electronic record varies with shift handover and is never measured. the present analysis treats that unmeasured lag as a design parameter rather than a nuisance and asks what benchmarking conclusions about artificial-intelligence sepsis onset prediction remain defensible when the lag is acknowledged. The analysis is a replication-based evaluation of the before-after-control-impact estimator under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power estimated across 2000 design cells at cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design required 54 clusters to detect a 5-percentage-point effect at 80% power. The results show that the intraclass correlation of 0.64 dominates the operating characteristics: at 8 clusters no effect size in the design grid reached 80% power, and the minimum detectable effect at 80% power fell steeply as clusters increased from 8 to 32. Benchmarking claims for sepsis early-warning models are conditional on a design that can absorb the lag, and it specifies the cluster counts and effect sizes at which such claims become defensible.