NabzSentry · Nabz familyPreprint
Fewer false alarms, with a limit you set on the real ones.
Most ICU VT alarms are false. Sentry silences them within an error budget you set.
74.8%
False alarms suppressed
At a 5% error budget
1.5%
Genuine alarms silenced
Expert-labelled VT alarms
0.953
AUROC
On par with best published
Try it
See Sentry triage an alarm.
Pick an alarm, run Sentry, then change the budget. Synthetic signals.
alarmBoth leads fast and wide; pulses fading.
Alarm · VTIllustrative
ECG 1
ECG 2
ABP
Pleth
Error budget
Illustrative ·suppress below 0.05 ·keep above 0.50
Triage this alarm
The decision
Three outcomes, because two forces a guess.
A two-way filter must commit on every alarm, even unreadable ones.
01
Keep
Likely real. It reaches the clinician as it does today.
02
Suppress
Likely false. Silenced, within the budget you set.
03
Defer
Too uncertain to call. A clinician reviews it.
Model4 waveform channels· ECG, arterial pressure, plethWorks with missing channels
The guarantee
You set the budget. Sentry certifies it.
5%
Error budget
Max genuine among suppressed
95%
Confidence
Checked on labelled alarms
3 of 3
Runs certified
At the 5% budget
If a budget can’t be certified, Sentry says so and estimates the labelled data it would need.
Research artifact, evaluated retrospectively. Not a medical device.