MCFANet: a multi-class fusion attention network for motor imagery EEG classification

This paper proposes a Multi-Class Fusion Attention Network (MCFANet) that combines the multi-class spatial filtering outputs of FBCSP with the spatiotemporal feature extraction capability of convolutional neural networks for multi-class motor imagery EEG classification. In multi-class motor imagery decoding, traditional spatial filtering methods extract effective discriminative spatial features but decompose the task into independent binary subproblems, and typically retain only energy statistics while discarding temporal dynamics. Deep learning methods can learn spatiotemporal features but must learn spatial patterns from the beginning, making it difficult to fully capture established neurophysiological priors under limited training samples.

MCFANet concatenates the spatial filtering outputs from all classes and sub-bands along the channel dimension to construct a virtual channel representation containing the discriminative responses of all classes. The full time series is preserved and fed into a convolutional module for spatiotemporal feature extraction, and a channel attention module adaptively reweights the feature maps to focus on the most discriminative representations. Four-class classification experiments were conducted on two public datasets.

On Dataset 2a, MCFANet achieved an accuracy of 67.94% ±13.70, outperforming FBEEGNet (63.98%) and EEGNet (58.79%). On the High Gamma Dataset, MCFANet achieved 87.10% ±10.09, improving over FBEEGNet by approximately 2.5 percentage points. Paired t-tests and effect size analysis confirm that the improvements over the main baseline methods are statistically significant.

The results suggest that reorganizing multi-class spatial discriminative responses into a unified representation that preserves temporal dynamics provides an effective path for bridging traditional spatial filtering and deep learning.