Deep adaptation network for subject-specific sleep stage classification based on a single-lead ECG. (May 2022)

• Record Type: Journal Article
• Authors: Tang, Minfang; Zhang, Zhiwei; He, Zhengling; Li, Weisong; Mou, Xiuying; Du, Lidong; Wang, Peng; Zhao, Zhan; Chen, Xianxiang; Li, Xiaoran; Chang, Hongbo; Fang, Zhen
• Is Part Of: Biomedical signal processing and control. Volume 75(2022)
• Publication Date: 2022-05
• Subjects: Domain adaptation network -- Generalization on subjects -- Sleep staging -- Single-lead ECG; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103548
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: An end-to-end domain adaptation network is proposed for subject-specific sleep staging. A single-lead Electrocardiograph (ECG) signal is utilized as the input. Multi-class focal loss and a domain aligning layer are combined to solve data imbalances and domain shifts. State-of-the-art performance of sleep staging based on single-lead ECG is obtained on the public SHHS2, SHHS1, and MESA datasets, respectively. The method has excellent performance for accuracy and Cohen's Kappa in terms of cross datasets. Abstract: Sleep plays a vital role in human physical and mental health. To accurately identify sleep structure under comfortable and convenient conditions, many machine-learning methods have been applied in the classification of sleep staging based on an Electrocardiogram (ECG); however, few works have solved the problem of generalization in the subject-specific sleep staging. The main reason for the problem is the difference of domain distribution across subjects. Most works are also classified based on ECG-derived signals and manual features. In this paper, we describe an end-to-end deep adaptation framework that classifies sleep stages into four classes based on a single-lead ECG to overcome the above problems. In particular, multi-class focal loss and a domain aligning layer based on the maximum mean discrepancy have been combined to solve data imbalances and domain shifts during three-step processing. We evaluate the method based on the three public datasets contained in SHHS2, SHHS1, and MESA. Compared to the performance of the model without domain aligning, the accuracy of the model for the public datasets has been improved by more than 20%, and the Kappa coefficient has also been improved by close to 0.4, which achieves state-of-the-art solutions compared to the baseline. In addition, the method has excellent performance for accuracy and Cohen's Kappa in terms of cross datasets. The proposed method, which confuses the domain-variant features, makes important contributions to the prediction of different subjects' sleep structures. The domain-adaptation setup, which varies across subjects and across environments, may provide a new approach to health monitoring.