Wearables can flag sleep apnoea but still cannot tell you how severe it is

Sleep apnoea, the condition in which airflow repeatedly ceases or decreases during sleep, often goes undiagnosed because the gold-standard test, polysomnography in a sleep laboratory, is expensive, inconvenient and in short supply. Wearable devices paired with artificial intelligence promise a very different route: convenient, affordable, objective and real-time monitoring worn on the body night after night. If they work well enough, they could flag the condition early and prompt timely intervention before complications develop. This review set out to establish how accurate that promise currently is, across detecting apnoea, distinguishing its type and grading its severity.

The reviewers searched six electronic databases for English-language research articles evaluating how well wearable AI identifies sleep apnoea, distinguishes its type and gauges its severity. Of 615 studies screened, 38 met the eligibility criteria. Two researchers independently selected the studies, extracted the data and assessed risk of bias using an adapted diagnostic accuracy tool. The evidence was then synthesised both narratively and statistically, with subgroup analyses probing which factors, such as algorithm choice or device placement, shaped performance.

Pooled accuracy for detecting apnoea events in breathing was 0.893, with sensitivity of 0.793 and specificity of 0.947. For detecting sleep apnoea itself, pooled accuracy was 0.869 with high sensitivity of 0.938 but weaker specificity of 0.752, which means a meaningful share of alerts will be false alarms. The clearest weakness was severity: accuracy for classifying severity level was just 0.651, although estimating the severity score itself, the Apnea-Hypopnea Index, fared better at 0.877. Performance varied with the type of algorithm, the type of data, the type of apnoea and where on the body the device was worn.

These results position wearable AI as a promising screening companion rather than a diagnostic replacement, echoing the wider story of consumer health technology: strong at flagging that something may be wrong, weaker at characterising exactly what and how badly. The authors recommend concurrent use with traditional assessment until better evidence supports reliability, and they call for certified commercial wearables, more research on central sleep apnoea, deep learning approaches and testing across different device placements. How these pooled figures translate to the specific watches and rings people actually own remains an open question.

If a wearable flags possible sleep apnoea, that signal is worth taking seriously, since pooled sensitivity for detection was high in these studies. But it is not a diagnosis, and it is definitely not a severity grade: the technology's accuracy at judging how bad apnoea is lagged well behind its ability to notice it at all. The weaker specificity also means some alerts will be false alarms. The sensible reading is that wearables act as an early-warning layer that still hands off to proper sleep assessment for anything that matters.