A convolutional neural network for automatic detection of sleep-breathing events using single-channel ECG signals

Abstract

Sleep apnea-hypopnea syndrome (SAHS) is a common sleep-breathing disorder, and the development of an automated method for detecting hypopnea and apnea is crucial for early prevention and treatment. Clinicians vary in their treatment approaches for sleep hypopnea and sleep apnea, often regarding sleep apnea as more severe. However, most previous methods for detecting sleep-breathing events were binary classifications that merged hypopnea and apnea into a single category, which are inadequate for auxiliary diagnosis and accurate sleep quality monitoring today. In this study, we proposed a convolutional neural network (CNN) model called TriGNet, which aims to detect hypopnea and apnea events. TriGNet utilizes single-channel electrocardiogram (ECG) signals from the MIT-BIH polysomnography dataset to learn and differentiate between normal, hypopnea, and apnea events. We designed a 2D feature extraction component in the model, which processes 1D ECG signal data as 2D to obtain additional feature information. In addition, this study proposed a dynamic data regulation mechanism for capturing subtle variations among the three categories of ECG signals. The TriGNet model can directly learn features from ECG signals and effectively classify sleep-breathing events. The proposed method achieved an accuracy of 94.08%, a macro F1 score of 93.02%, and a Cohen’s kappa coefficient of 89.91% on the test set. Experimental results show that the TriGNet model proposed in this study achieves state-of-the-art (SOTA) performance. You can find our source codes at https://github.com/MMMaoTS/TriGNet.