MSAF-Transformer: A Multi-scale attentive feature network for sleep apnea prediction based on single lead ECG signal

Sleep apnea (SA) is a sleep breathing disorder that can seriously affect the health of patients. It can result in hypertension, diabetes, and even sudden death. The ability to predict the occurrence of sleep apnea events in advance allows for timely treatment and intervention measures. In this paper, we propose a mutil-scale attentive feature network called MSAF-Transformer based on single lead electrocardiogram (ECG) signal. To enable the model to effectively capture both local details and global structure of the signal, we construct a multi-scale feature extraction network that captures information from different levels within the ECG signal. Additionally, in order to reduce network parameters and computational complexity, we have designed an attentive channel compression module which utilizes point by point convolution to decrease the number of feature channels. Furthermore, we employ a Squeeze-and-Excitation (SE) block to enhance feature expression ability and mitigate any negative impact caused by channel compression. Finally, the experimental results obtained from the publicly available Apnea-ECG dataset demonstrate the successful and efficient prediction of sleep apnea events by our network, thereby showcasing its potential for enhancing patient treatment in automatically regulated positive pressure (APAP) devices.