Obstructive Sleep Apnea Severity Classification using Sleep Breathing Sounds

Abstract

Obstructive sleep apnea (OSA) is a common sleep disorder. The symptom has a high prevalence and increases mortality as a risk factor for hypertension and stroke. Sleep disorders occur during sleep, making it difficult for patients to self-perceive themselves, and the actual diagnosis rate is low. Despite the existence of a standard sleep study called a polysomnography (PSG), it is difficult to diagnose the sleep disorders due to complicated test procedures and high medical cost burdens. Therefore, there is an increasing demand for an effective and rational screening test that can determine whether or not to undergo a PSG. In this thesis, we conducted three studies to classify the snoring sounds and OSA severity using only breathing sounds during sleep without additional biosensors. We first identified the classification possibility of snoring sounds related to sleep disorders using the features based on the cyclostationary analysis. Then, we classified the patients OSA severity with the features extracted using temporal and cyclostationary analysis from long-term sleep breathing sounds. Finally, the partial sleep sound extraction, and feature learning process using a convolutional neural network (CNN, or ConvNet) were applied to improve the efficiency and performance of previous snoring sound and OSA severity classification tasks. The sleep breathing sound analysis method using a CNN showed superior classification accuracy of more than 80% (average area under curve > 0.8) in multiclass snoring sounds and OSA severity classification tasks. The proposed analysis and classification method is expected to be used as a screening tool for improving the efficiency of PSG in the future customized healthcare service.