Fewer alarms, without silencing the one that matters.
Bedside monitors in intensive care produce arrhythmia alarms faster than clinical attention can absorb them, and most are false. The usual fix is a better classifier, which trades one error for the other. NabzSentry changes the shape of the decision instead, and puts the harmful error under a limit you set.
Three outcomes, because two forces a guess.
A binary alarm filter must commit on every event, including the ones where the signal has degraded past the point of being readable.
Keep
The evidence supports the alarm. It reaches the clinician exactly as it does today, with nothing added to the path.
Suppress
The evidence indicates a false alarm and the model is confident enough to act. This is the only decision that can cause harm, so it is the one placed under an explicit limit.
Defer
The evidence is too incomplete or too degraded to act on. A two-way classifier has to guess here. Sentry says so and hands the decision back.
You set the budget. The evidence has to earn it.
Suppressing an alarm is the only Sentry decision that can hurt a patient, so it carries an explicit limit. You choose the fraction of genuine alarms the system is permitted to silence. Sentry then holds to that limit with statistical confidence, established on a labelled sample and not on an assumption about how the model will generalise.
A limit like this costs something, and how much depends on how much labelled data you have. When the budget cannot be met, Sentry reports that plainly and says what would be needed to meet it.
NabzSentry is a research artifact. It is not connected to any clinical system, is not a medical device, and is not cleared for patient care. Evaluation to date is retrospective, on public de-identified data.
Alarm data to test this on?
We are looking for retrospective alarm cohorts and research collaborators. The full method and results publish with the paper.