Design: Systematic review and Bayesian bivariate diagnostic meta-analysis
Population: 13 studies, 9,983 adults
Primary outcome: Sensitivity and specificity for OSA against conventional sleep testing
Effect: Sensitivity 79.6% (55.5–93.8); specificity 76.5% (48.2–94.0)

A meta-analysis in the Journal of Medical Internet Research (25 September) pooled 13 studies, 9,983 adults, testing artificial intelligence models applied to photoplethysmography — the optical signal from pulse oximeters and many wearables — against conventional sleep testing.

Pooled sensitivity was 79.6% and specificity 76.5%, but the credible intervals were wide (roughly 55–94% and 48–94%). At higher severity thresholds specificity rose and sensitivity fell: for AHI of 30 or more, specificity was 85% and sensitivity 77%. Deep-learning models were more specific than older machine-learning approaches. Certainty was moderate.

Where polysomnography is scarce — including much of India — a cheap screening signal is attractive. But at this accuracy a negative result cannot exclude sleep apnoea, and a positive still needs confirmation. These tools may help prioritise who gets a sleep study, not replace it.

Do not use a wearable or oximeter-based AI result alone to rule out sleep apnoea.
Treat a positive result as a reason to arrange formal sleep testing.
Deep-learning models were more specific than older algorithms.
Clinical suspicion should still drive referral when screening is negative.

Why it matters

Consumer devices increasingly report sleep apnoea risk, and patients will bring those results to clinic.

The statistics, in plain English

A sensitivity of 80% means one in five people with sleep apnoea would be missed. The very wide credible intervals show the studies varied a lot, so real-world performance of any single device may be well above or below these averages.