Most bedside alarms in intensive care fire on physiology that is either an artifact or a real but self-correcting dip. Between 72% and 99% of alarms fall into these two groups, according to the systematic review of Chromik et al.. Signal processing already removes many artifacts. The open problem is the second group: telling a blood-pressure dip the body will correct by itself apart from one that will become a crisis unless someone treats it. We propose to build early-warning models that forecast two things. The first is whether a patient's blood pressure will fall. The second is whether the body's own compensating reflexes are still working. We would raise an alarm only when a predicted deterioration comes with signs that compensation is failing.
The compensation signals are deliberately simple, physiology-motivated features. Examples are whether heart rate still rises when pressure falls, the ratio of heart rate to systolic pressure, and pulse-pressure trends. They are cheap to compute at the bedside. The models would be judged the way a ward experiences them: under a fixed alarm budget measured in alarms per patient-hour, we would compare how many true events they catch, how early, and what share of alarms lead to treatment. In two sentences, the idea is this: predictive alarms should be ranked by whether the patient is losing the ability to self-correct, not just by how likely a drop is. Doing so should catch as many hypotension events while spending far fewer alarms on excursions that resolve without treatment.
Core Claim
At a matched alarm budget, a hypotension early-warning model that uses explicit compensation features and is trained to predict whether an excursion will need treatment will do two things better than a strong model trained only to predict the excursion. It will send a larger fraction of its alarms toward episodes that actually needed treatment. It will also lose no event sensitivity or lead time. We further claim that these features transfer across hospitals better than generic summary statistics. The point that separates this from prior work is the target and the evaluation. Prior predictive systems forecast whether deterioration occurs. We forecast whether it will self-resolve, and we report everything at a fixed alarm rate rather than as an F1 score.
Research Program
The setting is ICU hypotension, chosen because the strongest predictive alarm-fatigue result in the file lives there. Yoon et al. built a two-stage random forest on MIMIC-III. The first stage scores minute-by-minute vital-sign features: rolling statistics and Fourier spectra. The second stage models the trajectory of those risk scores. They added a 15-minute lockout that silences repeat alarms after an alert. Yoon et al. motivate this work, because they show physiological signatures diverge up to three hours before onset and that lockouts cut alarm volume roughly tenfold. They also constrain it, because they set a demanding operating point.
Yoon et al. modeled only the first hypotension episode, so the harder regime where fluids or vasopressors have already changed the dynamics is untested. They also did not separate excursions that needed treatment from those that did not. That is exactly the gap Chromik et al. call "actionability."
We will build three pieces.
The first piece is a set of labels. We label every excursion in a stay, not only the first. An excursion counts as actionable if it is followed within 60 minutes by a vasopressor start or fluid bolus, or if it persists past a fixed duration. It counts as self-resolving if it returns to baseline untreated. We will report results under both the intervention-based and persistence-based definitions. The reason is that treatment labels encode clinician habits, and a treated episode may have been one that would have resolved anyway.
The second piece is a small, fixed bank of compensation concepts computed over rolling windows. These are the sign and strength of heart-rate–pressure coupling, shock index and its slope, pulse-pressure trend, and respiratory-rate–SpO2 coupling.
These follow the domain-knowledge line represented by MINA, which steers ECG attention using signal derivatives and variance. They also follow the Wide and Deep Transformer, which won the PhysioNet 2020 Challenge with 22 handcrafted clinical features in a side branch next to a deep encoder. Unlike both, which classify the current recording, our concepts serve as inputs to a forward-looking prediction of self-resolution.
We keep the concepts cheap on purpose. Goodfellow et al. show the cost of expensive domain pipelines: their model timed out on the hidden PhysioNet 2020 test set and received a default score of −0.406. The same paper found that standard filtering hurt deep models, so we compute concepts on lightly cleaned signals and test robustness with added noise.
The third piece is the alarm policy. Each model produces two scores: $p_t$, the probability of an excursion within horizon $h$, and $q_t$, the probability that such an excursion would be actionable. The alarm score is $s_t = p_t q_t$. An alarm fires when $s_t$ exceeds a threshold $\tau$, followed by a lockout. We choose $\tau$ separately for each method so that
$$\frac{\#\text{alarms}}{\#\text{patient-hours}} \le B,$$
where $B$ is the alarm budget. Our headline $B$ is 0.79, Yoon et al.'s reported operating point. All comparisons are then curves of event sensitivity and actionable-alarm fraction plotted against $B$. This plays the role that Krijthe et al. gave to partial AUC restricted to low false-positive rates, and that the national-cost sepsis study gave to thresholds tuned per diagnostic category. In our case the budget is stated in the unit nurses actually experience.
The model class is kept separate from the claim. We fit the two heads both in a gradient-boosted tree and in a single-backbone sequence forecaster. If the gain shows up in only one model class, that tells us the source of the gain is the model rather than the concepts.
Baselines
The naive reference is a standard threshold alarm, for example mean arterial pressure below 65 sustained briefly, with and without the 15-minute lockout. It shows how many alarms a ward gets today and what any model must improve on.
The strongest existing competitor is a reimplementation of the stacked random forest of Yoon et al. with its 15-minute lockout. Its reported numbers are an AUROC of 0.88 and an AUPRC of 0.77 at 60 minutes before onset, 92.4% event capture, and about 16 alerts per hour across a simulated 20-bed ICU (0.79 per patient-hour). It is the strongest competitor because it already combines prediction, trajectory modeling and alarm suppression on the same data. We will first reproduce it on their first-episode cohort to check that our reimplementation lands near the reported values, and then extend it to all episodes. It gets the same features, the same data splits, the same hyperparameter-search budget, and the same budget-matched thresholding. Beating it means higher actionable-alarm fraction at $B = 0.79$ with event sensitivity no more than one percentage point lower and median lead time no shorter. This must hold on the all-episodes cohort, with confidence intervals from bootstrapping over patients.
The strongest deep predictive anomaly detector is A2P. It injects synthetic anomaly prompts so that a forecaster does not smooth away rare precursors, and it reports an F1 of 74.63 versus 51.44 for the next baseline on the MBA ECG benchmark. A second option is SC-JEPA, which forecasts in a latent space through a codebook of discrete regimes and reports F1 of 61.61 on PSM. We found no reported result on Lacuna for either method on MIMIC-III hypotension. We will train A2P on the same windows with its anomaly score thresholded to the same budget, with lockout. Beating it means the same criterion as above.
The personalized-threshold baseline follows Poole and Shah, who predicted patient-specific heart-rate distributions from EHR covariates. It is reported to suppress 44–50% of alarms, but at only 62% sensitivity against escalation markers. We adapt it to pressure.
The key ablations isolate our ingredient. The first, actionability-only, keeps the second head trained on actionability labels but gives it only Yoon's generic features. The second, concepts-only, adds the compensation features to the excursion head without the actionability target. If our full model does not beat both, the claim fails.
Why not just…?
Why not just keep Yoon's random forest and add the compensation features? This is close to our proposal and might be most of the gain. The concepts-only ablation tests it directly. If adding the features to Yoon's model matches the full system at every budget, then the predictive target adds nothing, and we will report that the contribution is the features alone.
Why not just train a classifier to predict intervention directly, with no special features? This is the most serious threat. Treatment labels already encode actionability, and a flexible model might discover coupling patterns on its own. The actionability-only ablation answers it within one database. The cross-database transfer test below is where we expect explicit concepts to matter, because learned proxies tend to latch onto hospital-specific charting habits. If actionability-only matches us both within and across databases, the concept bank is unnecessary, and we will say so.
Why not just lengthen the lockout or tighten the threshold on an existing detector? Lockouts cut repeat alarms for the same episode. Yoon et al.'s tenfold reduction came largely from this. But a lockout cannot tell a self-resolving first alarm from a dangerous one. Budget-matched curves make this visible: if lockout tuning alone reaches our actionable-alarm fraction at equal sensitivity, our second head is redundant. The evidence on post-hoc filtering is not encouraging. The Self-Beat Bank in the PhysioNet 2015 false-alarm study reached 0.960 specificity but only 0.116 sensitivity. Crude suppression tends to trade away exactly the events that matter.
Proposed Experiments
The primary experiment trains on MIMIC-III and reports budget curves on all excursions for both label definitions. It is done separately for first and subsequent episodes, because the post-treatment regime is where Yoon's design is untested.
The decisive transfer check trains on MIMIC-III and tests without retraining on eICU and on HiRID, re-choosing only $\tau$ to meet the budget. Chromik et al. note that these databases are almost unstudied for alarm fatigue. We predict that the gap between the full model and the actionability-only ablation widens under transfer. If it does not, the concepts carry no portable physiology.
Two further checks probe the mechanism. The first is a stratified analysis of excursions where heart-rate–pressure coupling was intact versus lost before the event. If compensation features matter, the full model's advantage should come mainly from correctly holding back alarms in the intact-coupling group. The second is a noise-robustness test that adds synthetic artifacts to the input signals. Its purpose is to confirm that the concepts do not collapse when fiducial points are unreliable. Fiducial points are the landmark positions, such as beat peaks, that the features are computed from.
Finally, we report per-sample inference cost for every method, so that no gain is bought with a model too slow to run at the bedside.
References
- Computational approaches to alleviate alarm fatigue in intensive care medicine: A systematic literature review
- Prediction of hypotension events with physiologic vital sign signatures in the intensive care unit
- MINA: Multilevel Knowledge-Guided Attention for Modeling Electrocardiography Signals
- A Wide and Deep Transformer Neural Network for 12-Lead ECG Classification
- Rhythm Classification of 12-Lead ECGs Using Deep Neural Networks and Class-Activation Maps for Improved Explainability
- Development and validation of an early warning model for hospitalized COVID-19 patients: a multi-center retrospective cohort study
- Optimizing the Implementation of Clinical Predictive Models to Minimize National Costs: Sepsis Case Study
- When Will It Fail?: Anomaly to Prompt for Forecasting Future Anomalies in Time Series
- SC-JEPA: Stabilizing Latent Predictive Learning for Time-Series Anomaly Prediction
- Addressing vital sign alarm fatigue using personalized alarm thresholds
- False arrhythmia alarm reduction in the intensive care unit