Nabz · A Roshan AI model family
Clinical reasoning on the signals themselves.
Nabz is our model family for bedside monitoring in surgery and intensive care. Each model says what the data support, and when they don’t.
3,442
Surgical cases
NabzGraph
5,037
VT alarms
NabzSentry
2
Preprints
On arXiv
The models
Two models today, one shared commitment.
Preprint
NabzGraph
Patient graphs from bedside monitoring
Typed relations that hold up when clinicians act on the vitals, each linked to the events behind it.
Preprint
NabzSentry
Certified three-way alarm triage
Keep, suppress or defer. You set the share of suppressed alarms allowed to be genuine; Sentry certifies it.
What makes it a family
Three rules every Nabz model follows.
The signal is the source
Every model reads bedside monitor streams directly.
Clinicians in the loop
NabzGraph tests clinician decisions and patient responses separately; NabzSentry defers uncertain alarms.
Earned, not assumed
Relations must beat negative controls; alarm budgets must be certified.
Both models are research artifacts built on retrospective data, not validated for clinical use.