End-to-End Sleep Apnea Detection Using Single-Lead ECG Signal and 1-D Residual Neural Networks

Abstract

Purpose Sleep apnea causes heart rate variability (HRV). HRV can be detected from the electrocardiography (ECG) signal and descriptors of HRV during sleep have been shown to be useful predictors of sleep apnea. In this work, we study the use of raw ECG signal and deep one-dimensional residual neural network (1-D ResNet) for end-to-end sleep apnea detection. Methods Our method uses raw single-lead ECG signal as an input to a 1-D convolutional neural network (CNN) with residual connections, exploiting CNN’s ability to learn distinguishing signal characteristics directly from the ECG signal and thereby forgoing the need for human engineered signal processing, feature extraction, and feature selection. In addition, we use weighted cross-entropy loss to account for the imbalance of apnea and non-apnea segments in our dataset, Bayesian optimization for fine-tuning the network hyperparameters, and data from current and adjacent epochs for predicting the label of the current epoch. The final ECG-based apnea detection network is evaluated on a dataset of 70 overnight ECG recordings. Results The proposed method achieved an accuracy of 93.05% (AUC\u2009=\u20090.9819) in detecting sleep apnea segments when considering adjacent epochs, thus, outperforming several baseline techniques. Furthermore, the method achieved 100% accuracy in separating sleep apnea recordings from normal recordings. Conclusion Our simple yet robust approach to ECG-based apnea detection demonstrates high accuracy. It has the potential to improve detection and diagnosis of sleep apnea and improve quality of life and health outcomes for millions of people worldwide.