SnoreOxiNet: Non-contact Diagnosis of Nocturnal Hypoxemia Using Cross-Domain Acoustic Features

Weiyan Yi, Xiuping Yang, Li Xiao, Weiping Tu, Xiong Chen, Yuhong Yang, Xinhong Li, Jie Lin

Abstract: Nocturnal hypoxemia, as a complication of OSAHS, is a chronic condition with a high prevalence and significantly increased risk of cardiovascular disease. However, due to the cost and inconvenience of the clinical diagnostic method, the patients’ visiting rate is extremely low. Snoring sound has been shown to be strongly correlated with nocturnal hypoxemia. Therefore, we build a snoring sound dataset, as well as a classification model based on the fusion of waveform features and spectrogram features of snoring sound, which estimates nocturnal hypoxemia status from snoring. The experimental results show that our method has an accuracy of 89.44%, a sensitivity of 93.68%, and a specificity of 82.81% for identifying subjects with or without severe nocturnal hypoxemia, and an accuracy of 71.65%, a sensitivity of 77.98% and a specificity of 22.46% for categorizing the four nocturnal hypoxemia severities. Our study provides a low-cost and convenient alternative method for diagnosing nocturnal hypoxemia by intelligent analysis of snoring sound, which can be easily recorded using smart phone.