Automatic sleep staging method based on CNN-BiGRU and multi-head self-attentionmechanism(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2025年第4期
- Page:
- 496-504
- Research Field:
- 医学信号处理与医学仪器
- Publishing date:
Info
- Title:
- Automatic sleep staging method based on CNN-BiGRU and multi-head self-attentionmechanism
- Author(s):
- ZHANG Xiaoli1; 2; ZHANG Xizhen1; 2; LIN Dongmei3; CHEN Fuming2
- 1. School of Medical Information Engineering, Gansu University of Chinese Medicine, Lanzhou 730000, China; 2. Medical SecurityCenter, the 940th Hospital of Joint Logistics Support Force of Chinese People’s Liberation Army, Lanzhou 730050, China; 3. School ofElectrical and Information Engineering, Lanzhou University of Technology, Lanzhou 730050, China
- Keywords:
- sleep stage; class balance; residual network; bidirectional gated recurrent network
- PACS:
- R318
- DOI:
- 10.3969/j.issn.1005-202X.2025.04.011
- Abstract:
- The study aims to address the issues of class imbalance in sleep EEG data and gradient vanishing or explosionphenomena that may occur when deep networks extract more features. An improved adaptive synthetic sampling technique isfirstly employed to perform data augmentation on the minority classes of sleep EEG data. Subsequently, convolutional neuralnetworks and residual networks are utilized to learn data features, while a 3-layer bidirectional gated recurrent network isapplied to explore deep temporal information and establish correlations between different sleep stages, enabling automaticfeature learning and sleep cycle extraction. Finally, a multi-head self-attention mechanism is adopted to enhance the model’sfocus on critical parts of the sequence, thereby completing the classification of various sleep stages. Experimental resultsshow that according to the AASM sleep staging criteria, the automatic sleep staging model integrating CNN-BiGRU andmulti-head self attention achieves an overall accuracy of 90.77% and a Kappa coefficient of 0.88 on the Sleep-EDF-20dataset after data class balancing, with the precision of N1 stage reaching 87.1%. On the Sleep-EDFx dataset, the modelattains an MF1 score of 0.84 while maintaining a precision of 77.2% for N1 stage classification. These metrics demonstratesignificant improvements in performance as compared with CNN-BiGRU model tested on the original dataset. Whenbenchmarked against other related studies, the proposed architecture exhibits superior sleep stage classification accuracy.These findings collectively validate the effectiveness and generalization capability of the proposed method.
Last Update: 2025-04-30