Coincidence Vetoes for ICU Alarms with a Guaranteed Miss Rate

Bedside monitors in intensive care units sound alarms for life-threatening heart rhythms, and somewhere between 80% and 88.8% of those alarms are false. Most are caused by a loose electrode, patient movement, or muscle noise corrupting the electrocardiogram (ECG), the electrical trace of the heartbeat (False arrhythmia alarm reduction in the intensive care unit). Recent deep models filter many of these alarms. None of them can say how many real emergencies they will silence, and that number is the one a hospital needs before letting software mute an alarm. Rare-event physics faced the same problem decades ago. A neutrinoless double-beta decay or dark matter search must reject almost all background while proving it keeps nearly all signal. It does this with vetoes: an event is thrown out only if an independent detector shows positive evidence that it was background. Each veto's signal loss is measured on large calibration samples rather than on the handful of real signal events.

We propose to bring that discipline to ICU alarms. An alarm raised by the ECG is suppressed only when the mechanical channels give positive evidence that the heart is still beating normally. Those channels are arterial blood pressure (ABP) and the photoplethysmogram (PPG), the optical pulse signal from a fingertip sensor. The evidence counts only if the pressure and optical pulses keep the timing the patient's own heart produced in the minutes before the alarm. The evidence from each channel is combined as an explicit likelihood ratio. The suppression threshold is set so that a statistical upper bound on missed true alarms stays below a chosen level. If a mechanical channel is itself unreadable, it cannot veto anything, and the alarm sounds.

Core Claim

Our central claim is that safe alarm suppression has two requirements. The suppressing evidence must be independent of the channel that raised the alarm. Its behavior under the "nothing is wrong" hypothesis must be calibrated per patient, from the abundant unlabeled data recorded just before the alarm. A suppressor built this way can be given a miss-rate guarantee that end-to-end classifiers cannot easily offer. It can also run in closed form on bedside hardware.

The literature already holds a cautionary result that supports the independence requirement. The patient-specific "Self-Beat Bank" in False arrhythmia alarm reduction in the intensive care unit compared incoming ECG beats against the patient's earlier ECG beats. It reached 0.960 specificity, but its sensitivity collapsed to 0.116. We read this as a failure of design, not of personalization. The test used the very channel that triggered the alarm, so a corrupted channel and a truly abnormal one looked alike. Our design moves personalization onto the channels that did not trigger the alarm.

Relation to Prior Work

Three lines of work attack ICU false alarms on the PhysioNet/Computing in Cardiology 2015 Challenge data, and all of them motivate this proposal.

The rule-based "regular activity test" in False arrhythmia alarm reduction in the intensive care unit discards an ECG alarm when ABP or PPG look rhythmic. It is the closest ancestor of our veto. However, it uses fixed population thresholds, no timing constraint between electrical and mechanical events, and no calibrated probability, and it misses 3.6% of true alarms.

Learned multimodal models make up the second line. Single-modal and multi-modal false arrhythmia alarm reduction using attention-based convolutional and recurrent neural networks pools ECG, ABP and PPG through attention and recurrent layers, with losses that weight missed true alarms heavily. A contrastive learning approach for ICU false arrhythmia alarm reduction compares the alarm window against a pre-alarm baseline window inside a twin-network encoder. Both learn the cross-channel relationship implicitly. They choose operating points on a few hundred true alarms and give no bound on misses.

The third line is A Diffusion Model with Contrastive Learning for ICU False Arrhythmia Alarm Reduction. It reconstructs the expected waveform with a diffusion model, a generator that produces samples by iteratively denoising random noise, and rejects alarms with high reconstruction error. It is interpretable, but it is slow at inference and was evaluated only on two rhythm types.

The method itself comes from physics. Enhancing the significance of astrophysical events with multimessenger coincidences gives an exact Bayesian update that combines timing and localization coincidences across independent detectors, raising one candidate from 7.7% to 86% probability. We reuse that structure, with the alarm-type base rate playing the role of the prior. The paper also warns that its updates depend on assumed source models, which constrains our design. Our analogue of that dependence is the model of what a true arrhythmia does to perfusion, and we keep that model deliberately conservative.

Measurement of double beta decay of 100Mo to excited states in the NEMO 3 experiment requires all decay products to share a vertex within time-of-flight bounds. This motivates our hard timing window between the electrical event and the arriving pressure pulse; the physiology of that delay is reviewed in Wearable Continuous Blood Pressure Monitoring Devices Based on Pulse Wave Transit Time and Pulse Arrival Time: A Review. The surface-alpha cut in α-event characterization and rejection in point-contact HPGe detectors rejected all surface backgrounds above 3 mm radius while losing only 0.2% of bulk events, and that loss was measured on calibration sources. This supplies the evaluation philosophy we import: report signal sacrifice as a first-class, separately measured quantity.

Research Program

The system has three parts.

The background model. In the five minutes before each alarm, we fit a small per-patient model of the mechanical channels. It captures the distribution of beat-to-beat intervals in ABP and PPG and the patient's pulse arrival time, meaning the delay from clean ECG beats to the arrival of the pressure or optical pulse. It also captures the distribution of mechanical pulse amplitude and the delay between ABP and PPG pulses. During the alarm window, the "background" hypothesis F says the heart is still beating as before and only the ECG is corrupted. Under F, mechanical pulses keep arriving at the pre-alarm rhythm and amplitude. The ABP-to-PPG delay stays inside a window set by the patient's own history. Any clean QRS complexes that survive, the sharp spike of each heartbeat on the ECG, must be followed by a pulse within the patient's arrival-time window.

The signal model. The "true alarm" hypothesis T is written conservatively per alarm type. Asystole and ventricular fibrillation imply loss of organized pulsatility. Extreme bradycardia and tachycardia imply a mechanical rate shift. Ventricular tachycardia is the hard case, because pulses may persist, and here the signal model allows reduced amplitude and an altered rate.

Combination and threshold. For channels $c$ with features $x_c$, the posterior probability that an alarm is true is

$$P(T \mid x) = \frac{\pi L}{\pi L + (1-\pi)}, \qquad L = \prod_{c} \frac{p(x_c \mid T)}{p(x_c \mid F)}.$$

Here $\pi$ is the training-set fraction of true alarms for that alarm type. The product over channels assumes the channels' noise is conditionally independent given the hypothesis. We test that assumption directly and fall back to the single most informative channel where it fails.

An alarm is suppressed only when $P(T \mid x) < \tau$ and every contributing channel passes a signal-quality check. The threshold $\tau$ is the largest value for which a one-sided Clopper-Pearson upper bound on the miss rate, computed on held-out true alarms, stays below a target $\varepsilon$. Its scale is set by small-count statistics. With zero misses among the 294 true training alarms, the 95% bound is roughly 3/294 ≈ 1%. That pooled level is attainable, but per-type bounds for the six true ventricular fibrillation alarms are not, and we will report that limit rather than hide it.

A second, secondary veto. Borrowing the trailing-edge analysis of the germanium detectors, we add a parametric descriptor of the ECG signal itself. It measures slow exponential baseline relaxation and leading-edge slew rate, the kind of pattern a peeling gel electrode produces. Because it uses the triggering channel, it is allowed to strengthen a mechanical veto but never to veto alone. That restriction is the lesson of the Self-Beat Bank.

Everything runs in closed form on window statistics, so the latency gap left by iterative diffusion sampling does not arise.

Baselines

The naive reference point is to pass every alarm. This gives 100% sensitivity and 0% suppression. On the 2015 training set that would mean, for example, 252 of 341 ventricular tachycardia alarms sounding falsely. Every method is reported as suppression gained over this reference at the realized miss rate.

The rule-based reference is the regular activity test of False arrhythmia alarm reduction in the intensive care unit. It reports sensitivity 0.908, specificity 0.838 and a Challenge score of 0.756, with a 3.6% false-negative rate. It is the most direct competitor because it uses the same idea of a mechanical cross-check. We reimplement it on our patient-level folds. Beating it means equal or higher specificity with a miss-rate upper bound below 3.6%.

The strongest learned competitor is A contrastive learning approach for ICU false arrhythmia alarm reduction. It reports a Challenge score of 84.47, 96% sensitivity and 86% false-alarm filtering. The attention model of Single-modal and multi-modal false arrhythmia alarm reduction using attention-based convolutional and recurrent neural networks reports 93.88% sensitivity and 92.05% specificity, and the diffusion model reports a ventricular tachycardia score of 74.10. These numbers come from different splits and score scalings, so we will not compare against them directly. We retrain the contrastive model and the attention model on identical patient-level folds, with the same hyperparameter search budget as our own tuning of window lengths and quality thresholds. We then compare all methods at matched realized sensitivity.

Our claim is not a higher Challenge score. Beating these models means removing at least as many false alarms at a certified pooled miss bound of about 1%. It also means degrading less when evaluation shifts away from the training conditions.

The key ablations isolate the proposal's ingredients. The first replaces the per-patient pre-alarm background model with population-level statistics. The second removes the hard arrival-time window. The third lets the ECG tail descriptor veto on its own, which predicts a Self-Beat-Bank-like loss of sensitivity.

Why not just…?

Why not just tune the rule-based cross-check's thresholds? Its 3.6% miss rate comes from fixed population thresholds that mistake weak but rhythmic pulsatility, or a wandering pressure baseline, for a beating heart. We include a version with tuned thresholds. If it matches our method at the same certified miss bound, then the per-patient background model adds nothing, and we will report that.

Why not just raise the threshold on the contrastive network until sensitivity is high enough? This is the real threat. If it suppresses as many false alarms at the same certified miss bound and holds up under shift, our method's remaining advantages are only latency and interpretability. We test it directly with three experiments. The first holds out an entire alarm type during training. The second evaluates with one mechanical channel removed. The third injects noise taken from false-alarm segments into true-alarm records, mimicking a calibration source. Our expectation is that the learned network's miss rate drifts upward under these shifts while the explicitly gated veto fails toward letting the alarm sound. That expectation is not yet evidence.

Why not just apply conformal calibration to a deep model's score? Conformal calibration certifies the miss rate just as our threshold rule does. What it cannot do is exploit unlabeled pre-alarm data to personalize the background hypothesis. We run a conformal-calibrated contrastive model as a baseline, and we also stack our likelihood ratio onto its score. If the stacked version does not beat both components, the coincidence evidence is redundant with what the network already learns.

Proposed Experiments

All experiments use the PhysioNet/CinC 2015 training cohort: 750 expert-annotated records with Lead II ECG, ABP and PPG across five alarm types. We use patient-level cross-validation because the 500-record hidden test set is not available to us. We report four things: the official Challenge score, sensitivity with its Clopper-Pearson upper bound on misses, suppression rate, and per-alarm latency on a microcontroller-class processor.

Two checks decide the proposal. The first is a matched-sensitivity comparison under the three shift conditions described above. The second tests whether channel evidence is in fact conditionally independent, by measuring cross-channel correlation of the features within each hypothesis. If independence fails badly, the multiplied likelihood overstates confidence, and the guarantee rests only on the empirical threshold.

A negative outcome would still be informative. That would be the case if the learned models match us at equal certified miss rates even under shift. It would show that the safety argument for physics-style vetoes in clinical monitoring rests on latency and auditability alone, not on accuracy.