An AI ICU early warning score predicts deterioration, such as circulatory failure, hours before it becomes obvious. The best models are accurate: one Swiss system predicted 90% of circulatory-failure events with 0.05 alarms per patient-hour. But ICUs already drown in alarms, so an AI score helps only if it replaces noise rather than adding to it.
Stand in an intensive care unit for ten minutes and count the sounds. Monitors chime for heart rhythms, blood pressure, oxygen levels, disconnected leads. Infusion pumps beep. Ventilators alarm. Phones ring. A nurse silences one alarm while walking to answer another. Much of it means nothing, and everyone knows it. That knowledge is exactly the danger.
Into this room, vendors and researchers are now bringing AI early warning systems: models that fuse vital signs, lab results and treatment data to predict circulatory failure, respiratory failure, sepsis or cardiac arrest hours in advance. Many of them are impressively accurate on paper.
The problem is that the ICU does not lack signals. It lacks attention. Adding more predictions to an environment already saturated with alarms can make care worse, not better, even when every individual prediction is correct. The ICU needs fewer, more actionable alerts. AI can deliver that, but only if hospitals treat it as a chance to subtract, not just to add.
How many alarms does an ICU patient generate?
The most detailed answer comes from a study at the University of California, San Francisco, published in PLoS One in 2014 by Barbara Drew and colleagues. The team captured every monitor alarm for 461 adult ICU patients over 31 days.
There were 2,558,760 alarms. Of those, 381,560 were audible, which worked out at 187 audible alarms per bed per day. Nurse scientists annotated 12,671 arrhythmia alarms in detail and found 88.8% were false positives. Even among the 168 true ventricular tachycardia alarms, 93% were not sustained long enough to need treatment.
The authors traced the problem to inappropriate alarm settings, persistent conditions such as atrial fibrillation that triggered repeated alerts, non-actionable events and algorithm weaknesses in the monitors themselves. The Joint Commission had already issued a sentinel event alert on alarm fatigue in 2013, recognising that clinicians exposed to constant false alarms become desensitised and can miss the ones that matter.
This is the baseline any AI early warning system enters. The question is not whether the model is accurate. It is what it does to that stream.
What is an AI early warning score in the ICU?
Traditional early warning scores, such as the UK’s National Early Warning Score, NEWS2, published by the Royal College of Physicians in 2017, add up points for abnormal vital signs. They are simple, transparent and widely used, especially on general wards to decide when to escalate care.
AI early warning systems do something more ambitious. They look at many variables at once, including trends over time, lab results and medications, and estimate the probability of a specific event within a specific window. Because they combine signals rather than reacting to each one separately, they can in principle be both earlier and quieter than threshold alarms.
The best-known demonstration came from Bern. In a study published in Nature Medicine in March 2020, Stephanie Hyland and colleagues trained a model on a high-resolution ICU database with 240 patient-years of data to predict circulatory failure. In the test set it predicted 90% of circulatory-failure events, 82% of them more than two hours in advance, with an area under the ROC curve of 0.94. The crucial number for alarm fatigue was this: the system raised on average 0.05 alarms per patient per hour, around one a day, far fewer false alarms than conventional threshold-based systems. It was also validated in an independent cohort.
That is what good ICU AI looks like: one integrated, earlier warning replacing many scattered ones.
Do AI early warning systems actually improve outcomes?
The strongest outcome evidence so far comes from outside the ICU, and it is instructive precisely because of how it was implemented.
Kaiser Permanente Northern California deployed its Advance Alert Monitor across 19 hospitals between 2016 and 2019. The model scored general-ward patients for risk of clinical deterioration. But alerts did not go straight to the bedside. They went to specially trained nurses working remotely, who reviewed the patient’s record and then contacted rapid-response teams when action was needed.
In a study published in the New England Journal of Medicine in November 2020, Gabriel Escobar and colleagues compared patients whose condition reached the alert threshold in hospitals where the system was live with similar patients in hospitals where it had not yet been deployed. Mortality within 30 days after an alert was lower in the intervention group, with an adjusted relative risk of 0.84.
The design is the lesson. Kaiser put a human filter between the model and the bedside. The remote nurses absorbed the alert burden, applied judgement, and passed on only what warranted a response. The frontline team received fewer, better-curated escalations. The model predicted; the programme decided.
Many ICU deployments do the opposite. The model’s output appears as another number in the record or another notification on a phone, with no one specifically responsible for acting on it and no reduction anywhere else.
Why can more predictions lead to worse ICU decisions?
Three mechanisms explain how a good model can make an ICU worse.
Attention is finite. Each alert competes with every other alert and with the patient in front of the clinician. A new channel of predictions does not come with new nurses.
Predictions are not decisions. Knowing that a patient has a 40% chance of circulatory failure in eight hours does not tell anyone what to do. Give fluids? Start a vasopressor? Order an echo? Watch closely? An alert without a recommended action forces the clinician to reconstruct the reasoning, which is slow and tiring.
Overlapping models conflict. ICUs increasingly run several scores: sepsis, deterioration, kidney injury, delirium risk. Each was validated alone. Nobody validated them together. When three scores disagree about the same patient, the clinician has to referee, and decision fatigue sets in.
This is a hospital-wide problem, not only an ICU one, and it echoes what we have argued about the EHR being where everyone else’s demands land. Every new tool is justified alone. The burden is carried in aggregate.
A worked example: a 20-bed ICU adds an AI score
Hypothetical, with illustrative numbers. A 20-bed ICU generates alarms at roughly the rate in the UCSF study, around 187 audible alarms per bed per day, or about 3,700 across the unit daily.
Option one: the ICU adds a circulatory-failure model with Bern-like performance, about 0.05 alarms per patient-hour, or roughly 24 alerts a day across the unit. It sends them to nurses’ phones. Nothing else changes. The new alerts are few and valuable, but they arrive as 24 more interruptions in a stream of 3,700, through a different channel, with no defined response. Some will be acted on promptly. Many will not.
Option two: the ICU uses the model to justify removing noise. It reviews default monitor settings, widens thresholds for non-actionable alarms where safe, suppresses repeat alarms for known persistent rhythms, and routes the AI alert to a charge nurse with a short response protocol. Suppose that cuts audible alarms by a third, about 1,200 fewer a day, while the 24 AI alerts carry a clear owner and action.
The algorithm is identical. In option one, it is one more voice in a noisy room. In option two, it helps make the room quieter.
Isn’t more monitoring always safer for the sickest patients?
The strongest counterargument is that ICU patients are the sickest in the hospital and deteriorate fast. Alarms exist because patients die when changes are missed. Suppressing or widening threshold alarms to make room for an AI score trades a known safety net for a model that was validated retrospectively and may perform differently in your unit. If a patient dies because a threshold alarm was switched off in favour of an AI prediction that did not fire, no one will be comforted by the model’s AUROC.
That concern is legitimate, and it sets the terms for doing this safely rather than a reason not to do it. Hard alarms for crisis values, such as asystole or severe desaturation, should stay. Changes to thresholds should be made on evidence from the unit’s own alarm data, not vendor defaults. AI scores should run silently in parallel first, so the unit can see what they would have caught and missed before relying on them. But the status quo is not safe either. Drew’s data show that a system in which almost nine in ten arrhythmia alarms are false is already failing patients, just invisibly, through desensitised staff. And staff themselves are finite: alarm fatigue is part of why ICU nurses leave, a problem we framed in the nurse shortage is a leak, not a drought.
What this means for ICU leaders, clinicians and vendors
For ICU and hospital leaders, create an alert budget. Track total alarms and alerts per patient-day, and the fraction that led to action. Require that any new predictive tool come with a plan for what it replaces, who owns its alerts and what the expected response is. Measure the unit’s alarm burden before and after, not just the model’s accuracy.
For clinicians, insist on alerts that come with context: what drove the score, how it has changed, and what the model’s typical lead time is. A number without a reason is harder to act on than a threshold alarm.
For vendors, the competitive advantage is not a higher AUROC. It is fewer, better-timed alerts with clear actions, integration into existing monitoring rather than a separate channel, and evidence from prospective deployments. Publishing alarms per patient-hour, as the Bern team did, should be standard.
For regulators and professional bodies, guidance on combining multiple predictive tools in one unit is overdue. Each model’s clearance says nothing about how it behaves alongside five others.
For patients and families, the practical point is that noise is not neglect. Asking the care team how they decide which alarms need attention is a reasonable question, and reporting any change you notice in a loved one is always worthwhile.
Frequently asked questions
What is alarm fatigue in the ICU?
Alarm fatigue is the desensitisation that occurs when clinicians are exposed to large numbers of alarms, most of which are false or non-actionable. It can lead to slower responses or missed critical alarms. A UCSF study found 187 audible alarms per ICU bed per day, and 88.8% of annotated arrhythmia alarms were false positives.
What is the NEWS2 score?
NEWS2, the National Early Warning Score 2, is a system published by the UK’s Royal College of Physicians in 2017. It assigns points for respiratory rate, oxygen saturation, blood pressure, pulse, level of consciousness and temperature, with the total guiding escalation of care. It is simple and transparent, but less tailored than machine-learning models.
How accurate are AI early warning systems in intensive care?
The best published models are highly accurate in retrospective and external testing. A 2020 Nature Medicine study predicted 90% of circulatory-failure events, 82% more than two hours in advance, with an AUROC of 0.94. Accuracy in validation does not guarantee benefit at the bedside, which depends on how alerts are delivered and acted on.
Have AI early warning systems been shown to reduce deaths?
Some evidence suggests they can, when paired with a structured response. Kaiser Permanente’s Advance Alert Monitor, used on general wards with remote nurse review and rapid-response teams, was associated with lower 30-day mortality after an alert, an adjusted relative risk of 0.84, in a 2020 NEJM study. It was not a randomised trial.
Will AI replace ICU monitors and alarms?
Not soon. AI scores run on top of existing monitoring data, and hard alarms for immediately dangerous events will remain essential. The more realistic role is to integrate many signals into fewer, earlier and more meaningful alerts, allowing units to reduce non-actionable alarms safely, based on their own data.
The most valuable thing an AI early warning system can do in intensive care is not to speak up. It is to let a dozen other alarms stay quiet.
Related on the GHC blog: Paying for Outcomes Rents Behaviour. It Doesn’t Buy It. and What Health Ministries Get Wrong About Airline Safety Culture.
How many alarms does your unit hear in a shift, and has any new tool ever made that number go down?
