0493 A Predictive Model Of Sleep Disordered Breathing Based On Deep Neural Network

Breathing sounds during sleep are disturbed in patients with sleep disordered breathing (SDB) and characterized by various acoustic properties. This study aimed to investigate acoustic biomarkers that can be used to detect the severity of SDB by analyzing the breathing sound collected from a large number of subjects during entire overnight sleep. Patients who visited a sleep center because of snoring or cessation of breathing were enrolled and underwent an full-night polysomnography (PSG). Breathing sound was recorded using a microphone equipped in the PSG. Then, audio features were extracted and a group of features significantly different between the SDB severity groups were selected as a potential acoustic biomarker. To assess the validity of the acoustic biomarker, classification tasks are performed using a deep neural network technique. Depending on the apnea hypopnea index of the subjects, four-group classification task was performed. Using 5-fold cross validation, we achieved an accuracy of 73.3% and a specificity of 91.7% in the four-group classification. The experimental evaluation demonstrated that the models trained on the proposed acoustic biomarkers can be used to estimate the severity of SDB. The acoustic biomarker of a subject may be useful to accurately predict the SDB severity of a patient from the sleep breathing sound, without conducting attended full-night PSG. This study has an implication that any device with a microphone, a smartphone for instance, has a potential to be utilized outside the specialized facilities as a screening tool for detecting SDB. Support (If Any):