Certified ICU alarm triage: suppress false VT alerts under a risk budget

A medical safety/eval note: three-way retain/suppress/defer triage on VTaC with a finite-sample style bound on silenced true alarms, and a clear warning about multiplicity eating the certificate.

Primary source: arXiv:2609.12365: https://arxiv.org/abs/2609.12365

What’s new: On the VTaC benchmark, 71% of ventricular-tachycardia alarms are false, but suppressing a real one can delay recognition of a dangerous arrhythmia. Certified AI Triage of ICU Alarms (arXiv:2609.12365, submitted 11 Sep 2026) reframes alarm reduction as three-way triage (retain, suppress, or defer) and bounds the harmful decision: among suppressed alarms, the fraction that were genuine stays below a user-set budget with 95% confidence under i.i.d. event sampling (clustered analysis when alarms share a waveform is a sensitivity check).

On the official split, a 5% budget certifies in all three seeds: suppressing 74.8% of false alarms while silencing 1.5% of genuine ones, at AUROC 0.953 and Challenge Score 83.33, numerically comparable to the strongest of eleven published systems. The central methodological point is multiplicity cost: the correction charges for every candidate, so a finer policy grid can certify strictly less. Under held-out calibration, the declared 885-cell grid certifies 1 of 15 fold-runs; choosing the grid on a separate selection partition certifies 8. The authors project calibration volume per budget (uncertifiable budgets become a design parameter) and report that adding a learned reliability dimension did not sharpen the certified frontier.

Why it matters: ICU alarm fatigue is a real clinical systems problem, and “higher AUROC” is not a safety story unless you say what happens to the true alarms you mute. Framing suppress as a budgeted risk with an explicit certificate is the right shape for safety-minded medical ML, and the multiplicity result is a useful corrective for anyone shopping policies on a huge grid.

Caveats: Preprint. The guarantee assumes i.i.d. event sampling; waveform clustering is only a sensitivity check. VTaC is a specific VT-alarm benchmark, not a full multi-vendor ICU monitor fleet. Challenge Score parity with prior systems does not equal regulatory clearance. The 885-cell failure under held-out calibration is the authors’ own warning: certification is fragile to how you select and charge for candidates. Treat this as a careful prototype of how to certify alarm suppression under a false-silence budget, not as a green light to auto-mute monitors.