Classification of High-risk Sleep Apnea with Arrhythmia Using Heterogeneous Ensemble Learning from Wearable Accelerometer

Sleep apnea poses a greater risk in arrhythmia patients. However, most sleep apnea studies have focused on healthy or sleep-disordered breathing subjects. In addition, it is not practical to use in everyday life by utilizing complex home-based portable devices that can obtain various biomedical signals for the classification of sleep apnea. In this study, a classification framework for sleep apnea was proposed for arrhythmia patients who are at high risk for sleep apnea. In particular, only accelerometers that can consider breathing patterns related to sleep apnea were used. In accelerometer signals, the normal breathing pattern of high-risk groups often resembles the sleep apnea pattern. To improve the classification performance for high-risk sleep apnea, this similar pattern was defined as the gray-zone, and a clear normal breathing and sleep apnea pattern was defined as the ideal group. We used a novel heterogeneous framework that combines (i) the feature generation method of statistics, relationship, and crest factor and (ii) the stacking ensemble method of support vector machine, multilayer perceptron, and logistic regression to classify sleep apnea pattern between ideal and gray-zone group. As a result, the proposed method obtained 0.71 ± 0.02 and 0.71 ± 0.01 of the F1-score in the ideal and grayzone group, respectively. The overall performance was higher than that of baseline models. The proposed ensemble learning framework could potentially be embedded in wearable devices to provide sleep quality assessment services to high-risk sleep apnea patients anytime, anywhere.