Automated Detection of Sleep Apnea Using Machine Learning: A Novel Approach Using Smartphone and Microphone for Breathing Sound Analysis

In this study, we evaluate the accuracy of a novel setup in the detection of apneas and hypopneas and estimating the apnea-hypopnea index (AHI). The study device setup consists of a microphone placed underneath the nose and a smartphone to collect the data. We recruited patients who were referred to the St. Josephs Hamilton Sleep Clinic for a sleep study. Data from our study device is collected simultaneously with polysomnography (PSG) in the sleep lab. A total of 26 patients were recruited, of which 2 dropped out during the data collection. Data from the microphone was too noisy for interpretation in 3 patients. Across the remaining 21 patients, the AHI based on their PSG ranged from 2 to 125 events/h, with an average AHI of 34 events/h. We used regression models trained on microphone audio data to identify noise and we developed an algorithm based on root-mean-square of the audio data for automatic detection of apneas and hypopneas. With reference to the PSG, our study device had a sensitivity of 94% and specificity of 87% in detecting apneas/hypopneas across a cumulative 120.9 hours of sleep data and more than 3700 such events. Our study device was able to accurately predict the AHI to within 4.3 events/h (+/-3.1). As such, we can conclude that our study device is accurate in identifying apneas/hypopneas, estimating the AHI and at ruling out cases of severe sleep apnea and can potentially be used as a screening test for this purpose. Our study device has the practical advantages of being very low cost and potentially more accessible as a screening tool, though further validation studies are needed to study accuracy across a larger population and for use at-home.