Dual-ChannelCVT: A Multi-Stage Combination of Convolutional Neural Network and Vision Transformer for Sleep Stage Classification

Sleep stage classification is essential for effective sleep monitoring and the diagnosis of sleep-related disorders. Traditional manual sleep staging is labor-intensive and requires substantial clinical expertise. To mitigate these challenges, automated sleep staging methods employing deep neural networks have attracted considerable interest. Given the continuous nature of sleep signals and the gradual transition between sleep stages, accurate classification requires modeling both local and global temporal dependencies. This study proposes a multistage deep learning framework that integrates convolutional neural networks with vision transformers (ViTs), further enhanced by residual connections, to capture discriminative features from electroencephalogram and electrooculogram signals. The model achieves strong performance, with an overall accuracy of 81.6%, a Cohen's Kappa coefficient of 74.75%, and an F1-score of 76.6%. To exploit both temporal and spectral characteristics of the signals, the time-domain data are transformed into timefrequency spectrograms using the continuous wavelet transform, which further improves classification performance. The model was evaluated on 153 full-night polysomnographic recordings, demonstrating robust and reliable performance in automatic sleep stage classification.