Personalized risk detection. Not population thresholds.
RPM Triage learns each patient's own normal, catches real deterioration earlier, and explains every decision in plain language — so your nurses trust it, not just tolerate it.
Built on adaptive per-patient baselines, transparent audit rationale, and human-reviewed calibration — never a black box.
Green — logLive triage readout
HR78 bpm
SpO297%
Resp16/min
Within this patient's own rolling baseline. No action needed — logged for trend review.
FHIR-formatted output
Stateless, device-key API
Deterministic audit trail
PhysioNet-validated
The problem
Static thresholds create the alarms nobody trusts
This isn't a hypothetical — alarm fatigue is one of the most documented, industry-wide patient-safety failure modes in monitored care, well before RPM Triage existed.
85–95%
of clinical monitor alarms, industry-wide, are estimated to require no clinical action at all.
80 deaths
industry-wide, from alarm-related adverse events reported to The Joint Commission's Sentinel Event database, 2009–2012.
1 threshold
applied to every patient in most legacy systems — regardless of age, condition, or that patient's own normal.
Industry-wide alarm-burden and sentinel-event figures per The Joint Commission's clinical alarm safety alert — not claims about RPM Triage specifically. The fixed-threshold description reflects standard static-threshold monitoring design generally.
How it works
Built to catch what population thresholds miss
Every reading is scored against that specific patient's own rolling baseline, dampened when context explains it, checked against a hard clinical safety floor nothing downstream can soften, then time-tracked so a transient spike never gets treated like sustained decompensation.
An adaptive mean + variance per vital, learned from that patient's own history — not a population norm.
02
Contextual filtering
Dampens — never deletes — deviations explained by exercise, circadian rhythm, or degraded device signal.
03
Hard safety overrides
Absolute crisis thresholds, checked first, every time — nothing downstream can bypass or soften them.
04
Persistence & digest
A transient spike stays a spike. Sustained elevation escalates. Everything else rolls into a daily/weekly digest.
The clearest proof
Same deviation. Different context. Different outcome.
Two patients, the identical heart-rate reading — one explained, one not.
Suppressed → Yellow, logged
HR 104 bpm
Recent activity · SpO2 stable at 97%
Elevated heart rate explained by recent activity — dampened, filed for trend review, never escalated.
Escalated → Red, urgent
HR 104 bpm
Resting, zero recent activity · sustained deviation
Same deviation, no contextual explanation, sustained at rest — escalated for immediate nurse review.
Patients stay in the loop
A lightweight app patients actually use
Every patient gets their own portal — installable straight from their browser, no app store required — to report how they're feeling and message their care team directly.
Tap-to-report symptoms, plain language
Direct messaging with their assigned nurse
Installs to the home screen, works offline
Push notifications when your team replies
Validation
Measured, not asserted
Every reading run through the same pipeline a real device would use — replayed against a real, public clinical dataset, not hand-written examples.
0%
Sensitivity on adjudicated adverse events
0%
Fewer urgent-tier alerts vs. static thresholds
0
Missed adverse events in validation
Measured against real ICU waveform data (PhysioNet's BIDMC dataset), comparing our RED_URGENT (immediate-page) alert volume against a fixed-threshold baseline — re-runnable against your own data. Some of that reduction is alerts eliminated entirely; some is alerts correctly downgraded to a non-urgent nurse-review queue rather than paging immediately.
Where we are today
These numbers come from bench validation against a real, public physiologic dataset (PhysioNet's BIDMC ICU waveform data) — not a live clinical deployment. This hasn't yet been run against a real RPM/CCM patient population in production.
Scope: this engine is built to catch acute vital-sign decompensation — a reading that's crossed a dangerous threshold or drifted meaningfully from a patient's own baseline, right now. It is not a sepsis-prediction or early-infection-detection tool, and we don't claim it is. We validated it against a real sepsis-onset dataset specifically to check that boundary, and it performed exactly as a vital-sign engine should on a task it wasn't built for: it caught the acute decompensation moments, not the earlier, subtler diagnostic criteria (labs, organ dysfunction) that define a clinical sepsis diagnosis before vitals become extreme. If you need early sepsis prediction, that's a different, purpose-built model — not this product.
Questions
Common questions
No. Every output is explicitly a Clinical Decision Support (CDS) recommendation for licensed-provider review — never a diagnosis and never a directive treatment instruction. Final clinical judgment always rests with the reviewing provider.
A disconnected device or an empty payload is routed to a device-status queue, explicitly labeled as a connectivity issue — never scored or presented as a clinical anomaly. Low-confidence or motion-degraded readings are down-weighted, not blindly trusted.
The baseline is a rolling window (weeks, not one reading). Before it's proven, the engine widens rather than ignores deviation, so a genuinely dangerous reading from a brand-new patient is never waved through for lack of history.
No. It prioritizes what a nurse looks at first and explains why in plain language — it never acts autonomously, and nurse feedback drives periodic, human-reviewed recalibration rather than the model silently updating itself.
Real, public ICU physiologic waveform data (PhysioNet's BIDMC dataset) — see the Validation section above, including the honest caveat that this isn't yet a live clinical deployment.
No, and we don't claim it does. This engine detects acute vital-sign decompensation — a dangerous threshold breach or a meaningful drift from a patient's own baseline, happening now. Early sepsis is often clinically diagnosed (via labs and organ-dysfunction criteria) well before vitals look extreme, which is a fundamentally different, purpose-built prediction problem. We validated against a real sepsis-onset dataset specifically to confirm that boundary, not to claim coverage we don't have.
Built to explain every decision
Every triage output comes with a plain-language rationale behind it — never a black box.