Staffable ICU early-warning with calibrated multi-centre validation and alert-budget guardrails

We present a staffable ICU early-warning system with calibrated multi-centre validation and alert-budget guardrails. A multi-horizon mortality model was developed on a large public ICU database and externally validated across independent hospitals at fixed post-admission landmarks (6/12/18/24 h). Evaluation was prespecified to emphasize actionability: discrimination, calibration (slope/intercept, Brier, ECE), decision-curve net benefit, and operating-point performance. To make deployment resource-aware, we translate threshold choice into alerts per 100 admissions and map alerts to clinician review time and net monetary benefit (NMB), yielding guardrails that bound workload while preserving utility. We further provide a silent-trial readout that estimates lead time prior to activation, supporting governance and cautious piloting. Across centres and horizons, the model maintained strong discrimination, tight calibration, and positive net benefit at clinically relevant thresholds. Guardrail selection produced workload-constrained operating points with stable subgroup performance. By coupling calibrated multicentre validation with alert-budget guardrails and deployment-proximal readouts, this study reframes ICU early-warning from maximizing AUC to delivering a staffable, value-aware program ready for monitored roll-out and periodic recalibration.