EEGCMCNet: A Hybrid Network for Sleep-Stage Classification

In this paper we introduce a hybrid neural network architecture designed for the analysis and detection of EEG signals, named EEGCMCNet (EEG CNN Mamba Capsule Network), which significantly optimizes the determination of sleep stages. EEGCMCNet integrates a multi-scale convolutional network (MSCNN), the selective state space series model (Mamba), the squeeze and excitation (SE) block and capsule networks. Within EEGCMCNet, the MSCNN initially extracts low and high frequency features from the EEG signals, then the features are fed into the Mamba module for dynamic time series analysis and feature integration. This is followed by the SE block that further optimizes channel weights, highlights important features, and suppresses secondary information. Finally, the capsule network performs detailed analysis of spatial relationships and executes precise classification. The strength of our model lies in providing a new solution to the neglected aspects in this field, that is, to accommodate local feature extraction, global context feature extraction for long time series data, and also focus on the spatial relationships between features of EEG signals. Experiments demonstrate that EEGCMCNet excels on three datasets, with multiple metrics for sleep stage classification surpassing existing advanced methods. This provides a new approach for deep learning analysis of EEG signals.