Interpretability of Shapley Attributions for Sepsis Alert Models Under Nurse-Initiated Vital Sign Backfill: A Computer Alert-Fatigue Evaluation

Sepsis alert models increasingly rely on Shapley attributions to justify individual alerts to bedside nurses, yet the reliability of those attributions under routine nurse-initiated vital sign backfill has not been characterised. the present analysis treats the attribution pipeline as a design problem rather than a modelling problem: when a nurse backfills a missing vital sign after an alert has fired, the Shapley value computed on the completed record may diverge from the value that would have been computed on the record as it stood at alert time, and the divergence is not random but structured by the backfill mechanism. A replication-based evaluation of the before-after-control-impact estimator is used to characterise the operating characteristics of the design 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. The design requires 54 clusters to detect a 5 percentage point effect at 80% power, and that the attribution error is bounded above by a quantity that grows with the backfill rate and the intraclass correlation. The contribution is a design-level criterion for when Shapley attributions can be trusted at the bedside, and a specification of the alert-fatigue consequences when they cannot.