ECG-based AI detects sleep apnoea by reading the autonomic response

Design

systematic review and bivariate random-effects diagnostic meta-analysis (PRISMA-DTA)

Population

13 studies of AI models detecting obstructive sleep apnoea from ECG-based physiological signals

Primary outcome

pooled segment-level sensitivity and specificity with summary ROC

Effect

sensitivity 0.88, specificity 0.89, AUC >0.90, moderate heterogeneity

Thirteen studies of artificial intelligence models detecting obstructive sleep apnoea from electrocardiographic and related physiological signals were pooled in a bivariate random-effects diagnostic meta-analysis. Pooled sensitivity was 0.88 and specificity 0.89 at segment level, with a summary ROC area above 0.90 and moderate between-study heterogeneity.

The more interesting claim is mechanistic. Performance was consistent across quite different model architectures, which the authors read as evidence that the models are locking onto stable physiological patterns rather than quirks of particular datasets. Their interpretation is that these systems are not detecting airway obstruction at all — they are detecting its downstream consequence, the autonomic response visible in cardiovascular signals.

That matters for how such a tool should be used and where it will fail. A model reading autonomic arousal will be most confident in patients whose apnoeas produce a brisk sympathetic response, and least reliable in those whose response is blunted — people on beta blockers, those with autonomic neuropathy from diabetes, the elderly. Segment-level accuracy is also not patient-level diagnosis: classifying 30-second epochs well does not establish that the derived apnoea-hypopnoea index matches polysomnography. As a triage tool ahead of a long waiting list it is promising; as a replacement for a sleep study it is not yet anything.

• Treat ECG-derived apnoea detection as triage, not diagnosis
• Expect reduced reliability in patients on beta blockers or with autonomic neuropathy
• Confirm with a sleep study before starting positive airway pressure
• Note that segment-level accuracy does not equate to a validated apnoea-hypopnoea index

Why it matters

If these models read autonomic arousal rather than obstruction, the patients they fail on are predictable.

Don't overread it

Segment-level diagnostic accuracy in research datasets — no study here shows patient-level agreement with polysomnography in routine practice.

The statistics, in plain English

Sensitivity 0.88 and specificity 0.89 are good figures, but they are pooled at segment level — the unit is a slice of recording, not a patient, so they do not tell you the chance a given person is correctly classified. An area under the curve above 0.90 describes discrimination across thresholds and says nothing about which threshold to use in practice. Moderate heterogeneity across only 13 studies means the pooled figure smooths over real differences in how the models were built and tested.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.