Sleep Apnea Severity Estimation From Tracheal Movements Using a Deep Learning Model

Objective: Sleep apnea is a chronic respiratory disorder and its standard assessment requires full night in-laboratory polysomnography (PSG). However, PSG is expensive, time-consuming, and inconvenient. Thus, there is a need to monitor sleep apnea with more convenient wearable devices. The objective of this study was to implement deep learning algorithms to monitor sleep apnea severity based on respiratory movements that can be easily recorded over the trachea. Methods: Adult individuals referred to the sleep laboratory at the Toronto Rehabilitation Institute for overnight sleep studies were included (N=69). Simultaneously with the PSG, an accelerometer was attached to the participant's suprasternal notch to record tracheal respiratory movements. Twenty-one features were extracted from the tracheal movements and used in a deep learning classifier to detect respiratory events. The apnea hypopnea index (AHI) was estimated as the number of events per hour of sleep. Results: The F1 score of the event-by-event detection algorithm was between 12% and 71% for different groups of sleep apnea severity. There was a strong correlation between the estimated and the PSG-derived AHI (r=0.86, p <; 0.0001). Using the AHI cut-off of 15, the sensitivity, specificity, and accuracy of diagnosing sleep apnea were 81%, 87%, and 84%, respectively. Conclusion: A combination of advanced machine learning algorithms and respiratory-related movements can accurately estimate sleep apnea severity and detect respiratory events during sleep. Significance: The proposed method can be implemented as a cost-effective and reliable wearable device for monitoring sleep apnea in the home and community.