Machine Learned Breath Sound Signatures for Obstructive Sleep Apnea

Resonea, Inc., Scottsdale AZ

Investigators

Linked publications & trials

Abstract

Abstract Sleep disordered breathing and in particular obstructive sleep apnea (OSA) is a major cause of workplace accidents, liability and days lost from work. Prevalence is rising with 30 million American sufferers alone. Despite this epidemic ? or perhaps because of it ? up to 40% of sufferers are undiagnosed and untreated. Lack of screening tools is a major bottleneck, since validated diagnostic tests of polysomnography (PSG) or home sleep testing (HST) have limited availability and are expensive and cumbersome. Simpler, widely applicable screening should be developed to identify individuals in the workplace at high-risk for SDB, for rapid triage to diagnostic testing and therapy. Our central scientific hypothesis is that OSA produces characteristic acoustic signatures, determined algorithmically in time- and frequency- domain analyses of ambient sound and by machine-learned networks, which can be used to screen for OSA. Our central business hypothesis is that such analysis should be possible using a consumer smartphone with no additional equipment. This work uses data available from our completed, IRB-approved study of 223 individuals at risk for OSA who underwent gold-standard sleep tests (PSG), providing multi-channel physiologic data on breathing health, and in parallel whole-night breath sound recordings from a smartphone. Aim 1 is a big-data approach to create an acoustic taxonomy, i.e. classification of breath sounds recorded by a smartphone, calibrated as normal, shallow, obstructed and absent using multichannel physiological data on the time-matched clinical PSG. Our team comprises trained sleep medicine physicians, occupational health providers, translational scientists and software developers. We will develop a library of acoustic signatures for loud gasping/choking/absent sounds characteristic of OSA and separate them from normal breath sounds. Aim 2 will develop a machine-learned respiratory index, using these acoustic signatures to track severity of OSA from concurrent all-night PSG, which predicts OSA severity in a separate validation cohort from our IRB-approved study. Supervised learning using convolutional neural networks will be used to derive the respiratory index. The significance of this proposal, scientifically, is that it will define digital acoustic signatures of disturbed from normal breaths and breathing patterns and, from an occupational health and business perspective, will use these tools to develop a widely accessible, rapid and efficient screening tool for sleep disordered breathing. Successful completion of this project may reduce the burden of workplace accidents from sleep disordered breathing. This project may shift the paradigm towards crowd-based screening and novel management strategies for this public health epidemic.