Accuracy of artificial intelligence-based prediction of moderate-to-severe obstructive sleep apnea using audio signals: A systematic review.

Obstructive sleep apnea (OSA) is a prevalent sleep-related breathing disorder that increases the risk of chronic diseases and negatively affects quality of life. Although polysomnography (PSG) remains the gold standard for diagnosis, its high cost, labor-intensive and time-consuming procedures highlight the need for alternative diagnostic tools. Advances in machine learning and consumer sleep technologies have accelerated the possibilities of early screening and risk assessments for various medical conditions, including OSA. The aim of this systematic review was to review the evidence for the detection of OSA via audio signals (smartphone, microphone) in adults. Of 122 papers identified, 15 were included. Whilst some studies demonstrated the trade-off between high sensitivity vs low specificity (or vice versa), six studies showed both high-to-very high sensitivity and specificity (≥80%). These results suggest detecting the risk of moderate-to-severe OSA from audio signals is plausible and sufficiently accurate. We review factors that may guide future research to increase the odds of training and testing algorithms to screen for OSA.