Abstract
Machine-learning sepsis early-warning systems generate alarm fatigue due to a high number of false alarms. The assumed solution: a second-stage classifier that filters out false alerts. We reproduce a gradient-boosted alarm, extract its stream at 2.78 alerts/bed-day, and train a suppressor for never-septic alerts, using a time-aware taxonomy to protect merely-early alerts, on the PhysioNet/Computing in Cardiology Challenge 2019 cohort (n = 40,336, two hospitals). It is highly discriminative (AUROC 0.855) but does not reduce alert burden relative to simply raising the threshold: on a matched frontier the cascade win at only 13.6% of operating points (median −6.9%). Suppressor confidence declines monotonically with primary risk (−0.43, replicated across both hospitals) confident on the low-risk false alerts a threshold already removes, near-chance on the high-risk alerts that drive fatigue. The signal is physiological, not care-process driven (ablating measurement-density features shifts AUROC by +0.0004); the failure transports across both hospitals. Discriminative performance is decoupled from suppression utility; predictable false alarms are not safely removable.