Classification and Characterization of Sleep Apnea using Machine Learning Methods on Sleep Studies

Abstract

Obstructive Sleep Apnea (OSA) Syndrome is a sleep disorder in which breathing rapidly starts and stops during sleep. It has been independently linked to multiple adverse health outcomes including an increased risk of hypertension, diabetes, cardiovascular disease, stroke risk, and overall mortality. OSA is diagnosed using an overnight sleep study measuring multiple high-resolution physiologic signals called a polysomnogram (PSG). PSGs are used to label sleep stage, or the cycles of sleep, and the occurrence of the apneas and hypopneas that define sleep apnea. The clinical diagnosis, prognostic determinations, and management of OSA depends the apnea-hypopnea index (AHI), a simple measure of the number of these events. Overall predictive value of AHI for OSA-related adverse outcomes is low and its clinical significance remains hard to quantify even after decades of sleep research. There is additionally a vast amount of data within PSGs that we can learn from and use to better quantify sleep apnea. Towards this goal, we develop models that efficiently and accurately characterize and measure sleep and respiratory events. We use deep learning models to classify sleep stage and detect apneas and hypopneas. We then develop models with engineered PSG-derived features for various cardiac outcomes related to sleep apnea using machine learning methods. The staging and event detection models allow the time-consuming and variable tasks in PSG scoring to be automated, and the development of new data-driven sleep-associated cardiac risk descriptors lays groundwork towards phenotyping sleep apnea. This work is significant in that it uses data-driven methods to lessen the burden of organizing and analyzing PSGs and provides new descriptors of data within PSGs that improve prediction of related cardiac outcomes.