A Mamba-Enhanced Spatial-Channel Recombination Temporal Network Using 4-D Feature Flow in EEG Signals

In recent years, extensive research efforts have been dedicated to the application of deep learning in EEG classification, yielding a plethora of proposed methods. Nonetheless, the majority of these methods have failed to comprehensively consider the information of EEG signals across the three essential dimensions, namely spatial, frequency band, and temporal domains, in a simultaneous manner. To this end, this paper proposes a Mamba-Enhanced Spatial-Channel Recombination Temporal Network (SCM-NET), using a four-dimensional EEG signal that has undergone DE cross-entropy as input. Firstly, the spatial-channel recombination module is exploited to recombine the spatial and frequency band features in accordance with the pre-configured activation threshold. This procedure effectively augments the spatial interactive feature information across different frequency bands. Subsequently, a lightweight convolutional neural network is harnessed to accentuate the most valuable information within the feature set. Finally, the Mamba network is applied to extract high-level temporal information features. We evaluated its performance using DEAP, MEEG, SEED, and SEED-VIG datasets and achieved recognized accuracy rates of 71.48%, 97.47%, 88.31%, and 99.89%, respectively, which outperformed the current state-of-the-art technology compared to existing methods. Codes are avaiable at https://github.com/NicoleRR2001/SCM-NET.