Detection of obstructive sleep apnea from single-channel ECG signals using a CNN-transformer architecture. (April 2023)

• Record Type: Journal Article
• Authors: Liu, Hang; Cui, Shaowei; Zhao, Xiaohui; Cong, Fengyu
• Is Part Of: Biomedical signal processing and control. Volume 82(2023)
• Publication Date: 2023-04
• Subjects: Obstructive sleep apnea -- ECG -- Transformer -- Deep learning; 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.2023.104581
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Obstructive sleep apnea (OSA) is a sleep breathing disorder that can seriously affect the health of patients. The manual diagnostic of OSA through the Polysomnography (PSG) recordings is time-consuming and tedious. Electrocardiogram (ECG) signals have been an alternative for OSA detection. This paper proposes a CNN-Transformer architecture for automatic OSA detection based on single-channel ECG signals. The proposed architecture has two fundamental parts. The first part has the aim of learning a feature representation from ECG signals by using the CNN. The second part consists mainly of the Transformer, a model structure built solely with self-attention mechanism, which is used to model the global temporal context and to perform classification tasks. The effectiveness of the proposed method was validated on Apnea-ECG dataset. The dataset consists of 70 ECG recordings with an annotation for each minute of each recording. The current and adjacent 1-min epochs were combined to form the 3-min input epoch. Besides, experiments were set up with different baseline deep learning models for sequence modeling to verify their effects on classification performance. The per-segment classification accuracy reached 88.2% and the area under the receiver operating characteristic curve (AUC) was 0.95. The per-recording classification accuracy reached 100% and the mean absolute error (MAE) was 4.33. Experimental results demonstrate that the Transformer structure and a 3-min input time window both effectively improve the classification performance. The proposed method can accurately detect OSA from single-channel ECG signals and provides a promising and reliable solution for home portable detection of OSA. Highlights: A novel deep learning architecture is proposed for OSA detection based on raw single-channel ECG signals. The CNN and Transformer structures are implemented to enhance the model performance. The current and adjacent 1-min epochs are combined to form the 3-min contextual input epoch. Achieving the state-of-the-art performance on the Apnea-ECG dataset.