Multi-task feature fusion network for Obstructive Sleep Apnea detection using single-lead ECG signal

Abstract

Efficient daily monitoring of Obstructive Sleep Apnea (OSA) and timely implementation of treatment is one of the important measures to ensure human health and sleep quality. Unfortunately, as the clinical gold standard, polysomnography (PSG) is complex to detect and cannot be popularized on a large scale. In recent years, OSA detection methods for single-lead electrocardiogram (ECG) signal based on deep learning have emerged. However, the detection performance of these deep learning methods based on a single learning task is often unsatisfactory. In order to solve this problem, we propose an OSA detection method based on one-dimensional multi-task feature fusion convolutional neural network (1D-MTFFNet). First of all, the borderline synthetic minority oversampling technique (Borderline-SMOTE) is used to alleviate the problem of data imbalance. Secondly, a multi-task learning method based on the combination of unsupervised feature learning and supervised feature learning is proposed to improve the feature extraction performance of the network. Finally, feature fusion between multiple tasks and between different levels is achieved by channel shuffling. We use the Apnea-ECG database of PhysioNet to conduct experiments. In terms of segment detection results, the proposed method achieves an accuracy of 91.13%, and the sensitivity and specificity reach 90.32% and 91.63% respectively. In terms of individual screening, the accuracy of the proposed method reached 100%. The overall performance is at the level of the most advanced methods.