An attention-based multi-resolution deep learning model for automatic A-phase detection of cyclic alternating pattern in sleep using single-channel EEG. (May 2023)

• Record Type: Journal Article
• Authors: Halder, Barproda; Anjum, Tanvir; Bhuiyan, Mohammed Imamul Hassan
• Is Part Of: Biomedical signal processing and control. Volume 83(2023)
• Publication Date: 2023-05
• Subjects: Cyclic alternating pattern -- Sleep EEG -- Deep learning -- Attention model; 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.2023.104730
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Sleep is a crucial part of human well-being. Many people suffer from various sleep disorders and insufficient sleep. The detection of the cyclic alternating pattern (CAP) of electroencephalogram (EEG) activity during sleep is essential for identifying and monitoring these problems. In this paper, we present a multi-resolution deep neural network model with temporal and channel attention for detecting A-phase and its subtypes. A multi-branch one-dimensional convolutional neural network (1D-CNN) is employed where each branch has different kernel sizes to extract features of different frequency resolutions automatically. An attention-based transformer network exploits the dynamic and temporal relationship between CAP event features extracted from the single-channel EEG data. Our model achieves 90.31% accuracy, 95.30% specificity, and 65.73% F1-Score in A-phase detection and 86.72% accuracy, 89.53% specificity, and 59.59% F1-Score in the detection of its subtypes, superior performance as compared to those of the recent approaches. Highlights: A multi-branch one-dimensional convolutional neural network (1D-CNN) is employed to extract features of different resolutions. Two types of attention mechanisms, temporal (multi-head attention), and channel attention, are used. Proposed method requires only single-channel EEG and minimal pre-processing. The model obtains 90.31% accuracy and 65.73% F1-Score in A-phase detection. In subtypes of A-phase detection, the model achieves 86.72% accuracy and 59.59% macro F1-Score.