AI for remote patient monitoring: signal, not noise
RPM programmes drown clinicians in alerts. How to build monitoring that surfaces deterioration without alarm fatigue.
Remote monitoring generates continuous data from people who are mostly fine. The clinical problem is not collection — it is that a programme producing hundreds of daily alerts trains staff to ignore all of them.
Thresholds are a bad model of a person
Fixed alert thresholds ignore that a reading which is unremarkable for one patient is a warning for another. Personalised baselines, built from a patient's own history, cut false alerts dramatically and catch the deteriorations that population thresholds miss.
Trend beats snapshot
A single out-of-range value is usually a measurement artefact — bad cuff placement, a device on the floor. Sustained direction of travel across several days is the clinically meaningful signal, and modelling that reduces noise more than any threshold tuning.
Threshold or baseline
Model of the patient
- Fixed thresholds
- A population average
- Personal baselines
- Their own history
False alerts
- Fixed thresholds
- High — normal variation trips them
- Personal baselines
- Far lower
Missed deterioration
- Fixed thresholds
- Patients whose normal sits inside range
- Personal baselines
- Caught, because the change is what matters
Clinician response
- Fixed thresholds
- Learns to ignore the queue
- Personal baselines
- Trusts and clears it
| Fixed thresholds | Personal baselines | |
|---|---|---|
| Model of the patient | A population average | Their own history |
| False alerts | High — normal variation trips them | Far lower |
| Missed deterioration | Patients whose normal sits inside range | Caught, because the change is what matters |
| Clinician response | Learns to ignore the queue | Trusts and clears it |
Missing data is data
A patient who stops taking readings is often the one you most want to hear about. Adherence monitoring belongs alongside physiological monitoring, and disengagement should raise its own gentle flag.
Monitoring that clinicians will actually use
Reduces alarm fatigue
- Baselines built from each patient's own history
- Trend over days, not a single out-of-range reading
- One prioritised queue, not a dashboard per metric
- Dismissal reasons fed back into thresholds
Causes it
- Fixed population thresholds for everyone
- Alerting on every artefact and cuff misplacement
- No signal when a patient stops taking readings
- Alerts with no visible reason attached
Design for the clinician's queue
The interface question is not 'is this patient abnormal' but 'who should I call first this morning'. Rank the queue, show why each patient is on it, and let a clinician clear it in minutes.
- One prioritised list, not a dashboard per metric
- The reason for each flag in one line, with the trend visible
- One-click acknowledge, escalate or dismiss with a reason
- Dismissal reasons fed back into the model's thresholds