Single-lead ECG based multiscale neural network for obstructive sleep apnea detection

Abstract

Obstructive sleep apnea (OSA) is a common sleep disorder characterized by frequent cessation of breathing during sleep, which cannot be easily diagnosed at the early stage due to the complexity and labor intensity of the polysomnography (PSG). Using a ECG device for OSA detection provides a convenient solution in the current Internet of Things scenario. However, previous intelligent analysis algorithms mainly rely on single scale network, therefore the discriminative ECG representations cannot be identified, which affects the accuracy of OSA detection. We report a multiscale neural network URNet for OSA detection by optimizing the deep learning networks and integrating Unet with ResNet. The URNet automatically extracts delicate features from the RR interval of single-lead ECG and processes convolution blocks with different scales by skip connections, so that the network can fuse features collected from both shallow and deep levels. For each OSA segment identification, URNet achieves an accuracy of 90.4%, a sensitivity of 83.3%, a specificity of 94.8% and an F1 of 89.6% on the Apnea-ECG dataset. The result indicates that our approach provides major improvements compared to the state-of-the-art methods. The URNet model proposed in this study for unobstructive OSA detection has good potential application in daily sleep health.