Automatic Obstructive Sleep Apnea Detection Based on Respiratory Parameters in Physiological Signals

Obstructive sleep apnea is a common sleep-disordered breathing disorder caused by repeated obstruction of the upper airway. Existing sleep apnea automatic detection models usually use the time-frequency domain and nonlinear features of physiological signals, and some are based on deep learning. The use of respiratory status parameters during sleep is mostly discussed for medical purposes, yet they are the most direct evidence of respiratory and pulmonary dysfunction. This paper proposed an automatic OSA detection method using respiratory parameters calculated from nasal airflow signals including respiratory cycle, respiratory rate (RR), tidal volume (TV), fractional inspiration (FIT), and minute ventilation (MV). Features chosen by statistical analysis were available to fed into machine learning models, and the results were compared. The selected features were further applied to ECG-derived respiration (EDR) signals to classify OSA and non-OSA patients. Among the models based on nasal airflow signal, extreme gradient boosting (XGBoost) had the best performance, with accuracy, sensitivity, precision, and F1 score of 86.42%, 83.82%, 87.55%, and 0.857, respectively. On the application of selected features in EDR signals, the results showed that the XGBoost model can achieve 82.76% accuracy and 85.97% precision, respectively. Our research can be used for screening and prognosis monitoring of sleep apneahypopnea syndrome patients with single-lead ECG sensors and provides a new path for the out-of-hospital management of sleep apnea patients.