Wearables and the IoMT
Continuous physiological sensing — PPG, SpO2, accelerometry — and the Internet of Medical Things: cleaning noisy streams, alerting on thresholds and doing it safely at the edge.
A wearable is a physiological laboratory strapped to a wrist. It measures heart rate from the changing colour of skin (photoplethysmography, PPG), estimates blood oxygen saturation from how much red versus infrared light is absorbed, and reads motion from an accelerometer. Do that every second for millions of people and you have the Internet of Medical Things — an enormous, noisy, personal data stream with real clinical potential.
The AI problem here is not image recognition. It is streaming signal processing: clean, segment, classify, alert — continuously, on a battery, under privacy and safety constraints.
Stream, clean, threshold, alert
At its core, a wearable monitor maintains a model of what "normal" looks like for this person and raises an alarm when the stream departs from it. Everything hard lies in the cleaning: motion corrupts the optical signal, contact is lost when the strap loosens, and the difference between artefact and arrhythmia is exactly the difference between a false alarm and a diagnosis.
Set thresholds and smoothing, and watch alert counts change on a synthetic wearable stream with genuine artefacts and a contact-loss gap.
Streams from a wearable: a tachycardia episode, a desaturation event, a motion artefact and a contact-loss gap. Set thresholds and smoothing, then count the alerts.
heart rate (alerts: 4)
SpO2 (alerts: 7)
skin temperature (alerts: 0)
- total alerts
- 11
- tachycardia
- 4
- desaturation
- 7
- invalid samples
- 5
Widen the smoothing window to suppress the motion artefact — and watch how it also blunts the sharp onset of the true tachycardia. That precision/robustness trade-off is the core design tension in ambulatory monitoring.
Illustrative synthetic physiology with genuine thresholding and smoothing. Real IoMT devices add sampling jitter, proprietary firmware filters, and regulatory-grade alarm logic; the basic pipeline of stream → clean → threshold → alert is unchanged.
Reading the signals
- PPG / heart rate. Light through tissue varies with blood volume, giving a pulse waveform. Heart rate is easy; rhythm analysis (atrial fibrillation) is subtle and exquisitely sensitive to motion artefact.
- SpO2. The ratio of red to infrared absorbance maps to oxygen saturation. Even small motion or poor contact shifts the estimate by several percent.
- Accelerometry. Provides step counts and activity, and also the signal quality needed to decide whether a PPG reading is trustworthy at all.
Edge or cloud
A wearable cannot stream raw multi-axis waveform data continuously; bandwidth and battery forbid it. So computation is pushed to the edge: filtering, segmentation and often a small classifier run on the device, and only events or summary features are transmitted. This is a real engineering trade-off between on-device accuracy (limited compute) and privacy/bandwidth (raw data is exactly what you would rather not upload).
Detecting arrhythmia
Modern ECG classifiers are 1-D convolutional networks — the same convolution machinery as audio and image models, with kernels sliding along time instead of space, producing a probability for each rhythm class. The landmark result is Hannun et al.'s single network matching cardiologist performance across a dozen rhythm classes on ambulatory single-lead ECG. The engineering that surrounds the model — resampling, filtering, handling variable-length windows, calibrating thresholds — is at least as important as the architecture.
Safety, privacy and regulation
Wearable health features live under medical-device regulation in many jurisdictions, and rightly so. A false negative can delay care; a false positive floods clinics with anxiety and cost. Continuous physiological data is among the most sensitive personal data that exists, so encryption, minimisation, consent and auditability are not optional.
Alarm fatigue is a safety failure
A monitor with a loose threshold and no artefact rejection emits so many false alarms that clinicians begin ignoring it — the failure mode known as alarm fatigue. Sensitivity, specificity and a clinically chosen operating point matter more here than raw accuracy, and every alert must be worth acting on.
Illustrative vs real
The widget generates synthetic heart rate, SpO2 and temperature with injected events and does real thresholding and moving-average smoothing. It does not model PPG optics, arrhythmia morphology or FDA clearance. Actual arrhythmia detection uses multi-lead ECG and 1-D CNNs validated prospectively; the stream–clean–threshold–alert skeleton shown here is faithful.
Check yourself
Eduspheria wiki · Applied AI, Medical AI
0 / 5 answered
From the exam paper
Modeled on NITJ AI-607, End-Sem May 2025
0 / 6 answered
Where next: the wearable's data is a signal, and the final chapter makes that explicit — sampling, Fourier spectra, filtering, and learned models over time.