Automated system for obstructive sleep apnea detection using heart rate variability and respiratory rate variability

Abstract

Listen

Translate article

Improving health care in rural areas is a major concern today. Sleep apnea is a common sleep disorder which often goes undiagnosed and leads to serious problems like stroke, heart attacks etc. Conventional diagnosis of sleep apnea is done by continuous recording of physiological signals for 6 to 7 hrs during sleep and then manually marking the events. This process is expensive and not affordable to people in rural areas. In this work we show that detection of obstructive sleep apnea (OSA) can be automated using machine learning techniques. We used heart rate variability (HRV) and respiratory rate variability (RRV) parameters derived from electrocardiogram (ECG) and respiratory effort signals (RES) respectively as input to the support vector machine (SVM) backend classifier. In a healthy person without apnea, there would be a rhythm in the autonomic nervous system (ANS). This rhythm will change in patients with OSA, even when no sleep disorder related events are clinically visible. We use HRV and RRV features as indicators of this rhythm. Based on this, our automated system for OSA detection (ASOSAD) detects OSA without waiting for any events to happen. We used the signals in rapid eye movement (REM) state of the sleep, as the chances of occurrence of apneas are more due to complete loss of muscle tone in REM. The REM state was detected using the chin EMG signal. 80% classification accuracy was obtained for the baseline system developed with time domain (TD) and frequency domain (FD) features of HRV and RRV as input. We found 10% absolute improvement in performance while using only the TD HRV features for developing the system. By analyzing the individual TD HRV and TD RRV based systems we found a complementarity in decisions. Subsequently We combined the decisions of the two systems by fusing the scores at the output of each system with weight which is calculated empirically using calibration data. For the fused system we obtained a performance improvement of 10% compared the best individual system, resulting in no error in OSA detection in REM stage. We are also exploring the possibility of developing a sleep stage independent system and the results will be reported in due course of time.