Automatic Sleep Staging for Patients with Abnormal Sleep Structure Based on Single-Channel EEG Signal Implemented by Convolution and Self-Attention Mechanism

• Published in: 2024 IEEE 25th China Conference on System Simulation Technology and its Application (CCSSTA) pp. 133 - 138
• Main Authors: Zhang, Qilong, Lu, Lin, Wu, Dongyue, Chen, Chao
• Format: Conference Proceeding
• Language: English
• Published: IEEE 21.07.2024
• Subjects: automatic sleep staging; Brain modeling; Decoding; Deep learning; Electroencephalography; Feature extraction; Investment; Multilayer perceptrons; Periodic structures; self-attention; Sleep; sleep structural abnormalities; Systems simulation
• DOI: 10.1109/CCSSTA62096.2024.10691731

With the aim of addressing the issue of the massive investment of human resources required for the current work of manual sleep staging in patients with structural abnormalities of sleep, we proposed an automatic sleep stage classification model based on single-channel electroencephalogram signals for the whole-night polysomnograms of 30 patients (16 males, 14 females) with sleep structural abnormalities in this study. The model extracts temporal-domain features from the raw single-channel electroencephalogram signal segments of each 30s epoch by using a convolutional neural network and a self-attention mechanism, and subsequently utilizes a multilayer perceptron to classify the features to accomplish the task of automatic 5 classification of sleep EEG signals. The proposed model achieves a mean classification accuracy of 81.46% and 82.92% on the C3-M2 and C4-M1 EEG channel, respectively. We also explored the effect of the self-attention mechanism on model performance enhancement and demonstrated the effectiveness of the self-attention mechanism for sleep electroencephalogram signal decoding. At the same time, this study provides evidence for the feasibility of deep learning models for the electroencephalogram decoding task in patients with sleep structural abnormalities.