Deep Learning Model for OSA Detection using Tracheal Breathing Sounds During Wakefulness

Abstract

The detection of Obstructive Sleep Apnea (OSA) during sleep is a simple and well-established technique; however, its detection during wakefulness is challenging. In this paper, we propose a deep learning model for the detection of OSA using only the tracheal breathing sounds spectrum as input. We employed our team’s previous dataset consisting of 109 subjects as non or mild-OSA with apnea/hypopnea index (AHI) < 15 and 90 subjects as OSA with AHI ≥ 15. All study subjects were referred to overnight polysomnography (PSG) to determine their AHI values. Tracheal breathing sounds were recorded in the supine position before proceeding to PSG while awake. The recording protocol was to have 5 deep breaths first through the mouth and then 5 deep breaths through the nose. Data were normalized and segmented into inspiratory/expiratory breathing phases; their power spectra were then calculated and fed to a deep learning model consisting of 71 layers. The results of 10 K-fold show that the proposed deep learning model achieved an accuracy of 74.9%, sensitivity of 76.1%, and specificity of 73.3%. Although these results are not as high as the previously reported analyses, they can be improved significantly by combining with anthropometric parameters and subgrouping subjects based on their age, weight, etc. This work aimed to show the potential of deep learning on this dataset despite the limited sample size. The results are encouraging to continue and improve the algorithm.