Fast and accurate screening of obstructive sleep apnea disorder by analyzing a few tracheal breathing sounds recorded while the individuals are awake
MetadataShow full item record
Obstructive sleep apnea (OSA) is one of the most common yet underdiagnosed sleep disorders. Undiagnosed OSA significantly increases perioperative morbidity and mortality for OSA patients undergoing surgery requiring full anesthesia. Polysomnography is the gold standard of diagnosing OSA, but it is expensive and time-consuming. Anthropometric risk factors such as sex, age, etc. are used in questionnaires to screen for OSA during wakefulness; however, they provide poor specificities (~20%) and no information about the disorder characteristics (e.g., total Arousal index, mean oxygen saturation (SpO2%), etc.). In this study, we present an objective and accurate tool along with a set of tracheal breathing sound characteristics with classification power for separating individuals with/without OSA and for predicting the disorder characteristics during wakefulness in a few minutes. Tracheal breathing sound signals were recorded during wakefulness in a supine position. Subjects were instructed to have a few deep breaths through the nose, then through the mouth. Study participants were referred to overnight polysomnography (PSG) assessment; their PSG reports were collected after their overnight-PSG study was completed. The signals were preprocessed; then, their power spectra and Bi-spectra were estimated. Different correlation and classification analyses were executed for predicting the OSA severity and the PSG parameters. The overall testing classification accuracy for AHI<5 and AHI>10 was 83.9%, and the odds ratio was 27.4. Results showed that different risk factors do affect the breathing sounds independent II of OSA severity. Using the least/most sensitive features to anthropometric features resulted in 72.1/83.6% testing classification accuracies (AHISupine =15). On the other hand, AwakeOSA algorithm resulted in 81.4% testing classification accuracy (AHI =15) on the blind-test dataset. In addition, this thesis investigated the predictability of PSG parameters from sound features. Using sound features and a Random-Forest classifier we achieved an unbiased classification accuracy up to 86.7% for predicting mean oxygen saturation, arousal index, and other PSG parameters. The results of this study show a superior OSA classification power of respiratory sound features compared to that of anthropometric features. Also, reducing the effects of the risk factors on the sound features resulted in a more reliable OSA screening algorithm. Finally, the results provide a new promising possibility for predicting the PSG parameters using the sound and anthropometric features without executing a full over-night PSG study.