ICU early-warning models now predict circulatory failure hours in advance, but raw predictions fire far too often to deploy. Most deployed systems therefore add a crude fix: after an alarm, they go silent for a fixed time. Yoon et al. use 15 minutes and Hyland et al. use 30 minutes. That timer does most of the work of cutting alarm volume. It also has two blind spots. It keeps an already-stable patient's alarm muted for the full window. Worse, it cannot re-alert on a patient who never recovers and keeps getting worse inside the window. We propose replacing the timer with a patient-specific recovery budget taken from engineering resilience metrics, which measure how deep, how long, and how fast a system falls below baseline and comes back. After an alarm, the system stays quiet while the patient's accumulated physiological deficit stays inside a learned "tolerable" envelope for the time elapsed. It re-arms as soon as the patient has truly recovered, and it re-alerts once the deficit exceeds what recovering patients typically accumulate.
The time-resilience literature supplies the tools. Schwind et al. define a system as $(k,l)$-resilient if its average cost over every window of length $k$ stays below $l$. Niu et al. give each subsystem a tolerance time (how long it can stay compromised) and a recovery time (how long it must stay healthy to rebuild its margin). The alarm-fatigue literature supplies the problem. Chromik et al. find that computational methods filter sensor artifacts well but cannot separate self-resolving true excursions from ones that need treatment. Both resilience papers assume automated control with perfect observation and say nothing about humans or noisy sensors. The clinical papers treat the post-alarm period as dead time. This proposal puts the resilience machinery exactly into that dead time.
Core Claim
A post-alarm silencing policy governed by a learned deficit-versus-elapsed-time budget should do better than fixed timers at a matched alarm rate. It should catch more deteriorations that begin or continue during what would have been a lockout, including repeat events after treatment. It should also cost no more alarms per patient-hour than the best tuned fixed timer. The contribution is the decision layer that sits on top of any risk model and decides whether to speak again. It is not a new risk model and not new input features, and the experiments are built to show that the policy structure, rather than extra information, is what matters.
The rule works as follows. Let $P(t)$ be mean arterial pressure smoothed with a Gaussian rolling mean, following Cadet et al., who use such means to separate real degradation speed from telemetry noise. Let $P_0$ be the patient's personal baseline, estimated from the preceding stable hours in the spirit of Poole and Shah. After an alarm at time $t_a$, the cumulative deficit is
$$D(t) = \int_{t_a}^{t} \max\big(0,\; P_0 - P(s)\big)\, ds,$$
which is the area of the degradation valley in the sense of Zeng et al.. The policy re-alerts at the first time $t$ at which
$$D(t) > B_{\theta}(t - t_a \mid x_{t_a}),$$
where $B_\theta$ is a budget curve. It is non-decreasing in elapsed time and conditioned on the patient's context $x_{t_a}$ at the alarm: vasopressor status, baseline, and risk score. It is learned so that trajectories exceeding it are enriched for progression to circulatory failure. The policy also re-arms early once $P(t) \ge P_0$ has held for a learned recovery time $\phi$, Niu's minimum healthy duration. The budget curve plays the role of Schwind's resiliency function, estimated from data instead of derived from a known model.
Research Program
We first build the event setting. We use the HiRID circulatory-failure task from the HiRID-ICU-Benchmark, at two-minute resolution. Unlike Yoon et al., who restrict evaluation to the first hypotension episode because treatment shifts baselines, we keep every onset. We tag each onset as occurring during vasopressor or fluid therapy or not, so repeat events after treatment become an explicit evaluation stratum rather than an exclusion. The base risk model is a LightGBM on summary features, reproducing the benchmark's strongest configuration. It is frozen, so every policy we compare sees identical risk scores.
Second, we learn the budget. Every post-alarm window in the training split is labeled by whether circulatory failure newly begins or persists within a fixed horizon. $B_\theta$ is fit as a monotone curve, using a small set of elapsed-time knots and quantile regression under a monotonicity constraint. Its conditioning uses only a few context variables. The target is one number, the alarm rate, which sets the operating point. $\phi$ is tuned jointly on validation data. A small parameter count is deliberate: the benchmark evidence that trees beat deep sequence models on this data argues against a high-capacity policy.
Third, we test transfer. The budget learned on HiRID is applied unchanged to MIMIC-III, where Hyland et al. report that circEWS-lite kept an AUROC of 0.939. This tests whether a structured, physiologically anchored rule survives the practice drift Hyland et al. flag.
Fourth, and briefly, clinicians review a stratified sample of alarm pairs retrospectively, blinded to which policy produced them. For each re-alert that our policy issues and the timer suppresses, and vice versa, reviewers judge whether it was actionable. Chromik et al. found that only 8 of 69 alarm-fatigue papers involved clinicians at all.
Baselines
The naive reference is the frozen risk model with a threshold and no silencing. It sets the ceiling on recall and the floor on alarm burden. Yoon et al. report that the unsilenced setting produced roughly 159 alerts per hour in a 20-bed ICU.
The strongest existing method is the circEWS silencing policy of Hyland et al.. It is a 30-minute silence that is broken only if the patient recovers above threshold and then deteriorates again. Hyland et al. report an AUROC of 0.94 and AUPRC of 0.63, and state that circEWS detects 90.0% of events at 0.05 alarms per patient per hour. On the benchmark split, LightGBM reaches an AUPRC of 38.9 per Yèche et al.. Their cohorts and definitions differ, so we reimplement the circEWS rule on our frozen scores rather than compare against the published figures. To keep the comparison fair, we also give the fixed timer the same tuning budget as our policy: a grid over timer lengths from 5 to 120 minutes, chosen on the same validation split.
We add two stronger simple policies that also use post-alarm information. The first is hysteresis: re-arm when the risk score drops below a lower threshold. The second is escalation on risk rise: re-alert during silence if the score increases by a tuned margin.
The key ablation keeps the same area-based deficit but replaces the learned context-conditioned budget with a single fixed area threshold. This isolates whether personalization over elapsed time matters.
Our policy counts as beating the baselines only if two things hold. First, at alarm rates matched to within 5% of each baseline's validation operating point, it achieves higher event recall. Second, it has a lower rate of "blackout misses", which are events whose onset or worsening falls inside a silence window, with the improvement concentrated in repeat events after treatment.
Why not just…?
Why not just tune the fixed lockout length? Tuning trades blackout misses against alarm volume along a single dial. Shortening the timer to catch non-recovering patients re-alerts on every stable patient too. Our Pareto curves of alarm rate against recall will show whether any timer length reaches the budget policy's frontier. If one does, the claim fails.
Why not just use hysteresis or re-alert when the risk score rises? This is the most serious threat. Both rules use post-alarm information and have one or two parameters. Our argument is that the risk score is trained to predict onset, not how a patient responds after treatment begins. The score may therefore stay flat or saturate while blood pressure fails to rebound under therapy. That is precisely the repeat-event stratum Yoon et al. excluded. The experiment settles the question directly. If escalation on risk rise matches our recall on post-treatment events at equal alarm rates, the resilience budget adds nothing, and we will report it that way.
Why not just add deficit-area and duration features to the LightGBM, or train a classifier that predicts whether a re-alert is needed? Features change the score, not the silencing rule, so even a better score still passes through a fixed timer. We test this by retraining the risk model with deficit features and applying the tuned timer. An unconstrained re-alert classifier on the same inputs is also a fair competitor. We will run it, and expect it to match in-distribution. The case for the monotone budget is transfer to MIMIC-III and auditability, and the transfer experiment decides whether that case holds.
Why not just personalize the thresholds? Poole and Shah suppress 50.5% of low and 44.1% of high heart-rate alarms. Their clinical-event evaluation, however, reaches only 62% sensitivity and 49% specificity. Their method also ignores how long an excursion lasts, which is the quantity our budget adds. We already borrow their personal baseline for $P_0$, so the ablation that removes elapsed-time conditioning covers this alternative.
Proposed Experiments
The primary experiment is on HiRID. It reports alarms per patient-hour, event recall, lead time, blackout-miss rate, and recall of repeat events after treatment, across all policies on one frozen LightGBM. It adds Pareto curves over each policy's full parameter range and bootstrap intervals over patients. The decisive check is the repeat-event stratum at the circEWS-matched alarm rate. There, the budget policy must beat the tuned timer, hysteresis, and escalation on risk rise, or the central claim is false.
The second experiment applies the HiRID-trained policy unchanged to MIMIC-III using mean arterial pressure. It checks whether its advantage over the unconstrained re-alert classifier grows under this shift.
The third is a noise stress test. Artifacts are injected into blood-pressure streams to see whether Gaussian smoothing keeps the deficit integral from triggering spurious re-alerts. This speaks to the perfect-observation assumption that Schwind et al. make explicit.
The last is the blinded clinician review. It checks that the alarms our policy adds are judged actionable, which the numbers alone cannot establish.
References
- Prediction of hypotension events with physiologic vital sign signatures in the intensive care unit
- Machine learning for early prediction of circulatory failure in the intensive care unit.
- Formalization of resilience for constraint-based dynamic systems
- A Compositional Resilience Index for Computationally Efficient Safety Analysis of Interconnected Systems
- Computational approaches to alleviate alarm fatigue in intensive care medicine: A systematic literature review
- Quantitative Resilience Modeling for Autonomous Cyber Defense
- Addressing vital sign alarm fatigue using personalized alarm thresholds
- FAIR: Facilitating Artificial Intelligence Resilience in Manufacturing Industrial Internet
- HiRID-ICU-Benchmark - A Comprehensive Machine Learning Benchmark on High-resolution ICU Data