SE-MSResNet: A lightweight squeeze-and-excitation multi-scaled ResNet with domain generalization for sleep apnea detection

• Published in: Neurocomputing (Amsterdam) Vol. 620; p. 129201
• Main Authors: Zhao, Yuxuan, He, Haitao, Wang, Qian, Yu, Lu, Ren, Jiadong
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2025
• Subjects: Attention; Domain generalization; Multi-scale; ResNet; Sleep apnea detection; Sleep apnea detection; Multi-scale; Attention; Domain generalization; ResNet
• ISSN: 0925-2312
• DOI: 10.1016/j.neucom.2024.129201

Sleep apnea (SA) is a common contributing factor to many sleep-related and cardiovascular diseases. The non-invasive sleep apnea monitoring method based on a single-lead Electrocardiogram (ECG) is essential for its prevention and treatment. Although previous attempts to detect SA have achieved high classification performance, two challenges still need to be addressed: (1) the non-stationary nature of ECG signals makes it difficult to effectively design a deep feature-extracting algorithm to extract typical sleep apnea-related features from ECG segments. (2) generalizing deep learning-based SA detection models in real clinical scenarios is challenging due to the individual differences among subjects. This paper presents a lightweight squeeze-and-excitation multi-scaled ResNet model named SE-MSResNet with domain generalization for SA detection. First, our work introduces a multi-scaled residual block (MSResBlock) to extract temporal features from RR intervals and R-peak amplitude sequences derived from single-channel ECG signals. Next, a squeeze-and-excitation attention module is applied to fuse the features extracted from the two above ECG-derived signals. Finally, we propose a jointly shared feature training strategy based on domain adversarial generalization and form a unified training framework. It aims to minimize the feature distribution discrepancy between the source domains (training subjects) and enhance common sleep apnea-related features, thus improving the proposed model’s generalization performance for unseen subjects. Experimental results on the publicly available dataset validate the effectiveness of our proposed model, surpassing the performance of state-of-the-art baselines.