A single-channel EEG based automatic sleep stage classification method leveraging deep one-dimensional convolutional neural network and hidden Markov model

• Published in: Biomedical signal processing and control Vol. 68; p. 102581
• Main Authors: Yang, Bufang, Zhu, Xilin, Liu, Yitian, Liu, Hongxing
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.07.2021
• Subjects: Convolutional neural network; Electroencephalogram; Hidden Markov model; Sleep stage classification; Subject-independent testing; Subject-independent testing; Hidden Markov model; Sleep stage classification; Electroencephalogram; Convolutional neural network
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2021.102581

•Our proposed 1D-CNN-HMM model combines 1D-CNN and HMM. 1D-CNN could extract features from raw EEG to perform epoch-wise classification, and HMM works as post-processing step to correct unreasonable sleep stage transitions.•We have demonstrated that HMM refinement is effective for 1D-CNN, and HMM improved the classification performance of 1D-CNN by improving the performance on S1 and REM stages with p < 0.05.•Results demonstrate that our method outperformed most of the existing single-channel EEG based methods under subject-independent paradigm using Sleep-EDFx and DRM-SUB datasets. Sleep stage classification is an essential process for analyzing sleep and diagnosing sleep related disorders. Sleep staging by visual inspection of expert is a labor-intensive task and prone to subjective errors. In this paper, we proposed a single-channel EEG based automatic sleep stage classification model, called 1D-CNN-HMM. Our 1D-CNN-HMM combines deep one-dimensional convolutional neural network (1D-CNN) and hidden Markov model (HMM). We leveraged 1D-CNN for epoch-wise classification and HMM for subject-wise classification. The main idea of 1D-CNN-HMM model is to utilize the advantage of 1D-CNN that can automatically extract features from raw EEG, and HMM that can utilize sleep stage transition prior information of adjacent EEG epochs. To the best of author's knowledge, this is the first implementation of 1D-CNN connected with HMM in automatic sleep staging task. We used Sleep-EDFx dataset and DRM-SUB dataset, and performed subject-independent testing for model evaluation. Experimental results illustrated the overall accuracy and kappa coefficient of 1D-CNN-HMM could achieve 83.98% and 0.78 on Fpz-Oz channel EEG from Sleep-EDFx dataset, and achieve 81.68% and 0.74 on Cz-A1 channel EEG from DRM-SUB dataset. The overall accuracy and kappa coefficient of 1D-CNN-HMM outperformed other existing methods both on two datasets. In addition, the per-class performance of 1D-CNN-HMM is significantly higher than 1D-CNN on S1 and REM sleep stages with p<0.05. Our 1D-CNN-HMM outperformed other existing methods both on two datasets. Results also indicated that HMM improved the classification performance of 1D-CNN by improving the performance on S1 and REM stages.