• Multi-Branch GAT-GRU-Transformer for Explainable EEG-Based Finger Motor Imagery Classification

Zhuozheng WangYunlong Wang^*

• Beijing University of Technology, Beijing, China

Electroencephalography (EEG) signals provide a non-invasive, real-time approach to decoding motor imagery (MI) tasks, such as finger movements, with significant potential for brain-computer interface (BCI) applications. However, the complexity, noise, and non-stationary nature of EEG signals challenge accurate classification. Traditional methods, such as Common Spatial Pattern (CSP) and Power Spectral Density (PSD), struggle to capture the multidimensional dynamics of EEG data due to their reliance on manually engineered features. In contrast, recent deep learning models, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), often focus on single-dimensional features and lack interpretability, which can limit their applicability in contexts requiring detailed neuroscientific understanding.This study proposes a novel multi-branch deep learning framework, termed "Multi-Branch GAT-GRU-Transformer," to address these issues. The model leverages parallel branches to extract spatial, temporal, and frequency features from EEG signals. Specifically, Graph Attention Networks (GAT) model spatial relationships, a hybrid Gated Recurrent Unit (GRU) and Transformer architecture captures temporal dependencies, and one-dimensional CNNs extract frequency characteristics. Feature fusion integrates these representations for robust MI classification. To enhance interpretability, SHAP (SHapley Additive exPlanations) and Phase Locking Value (PLV) analyses are incorporated, elucidating the neural basis of model decisions.Evaluated on the Kaya dataset, the model achieves a classification accuracy of 55.76% in a five-class MI task. Ablation studies validate the contribution of each component, while SHAP and PLV analyses reveal the model ' s reliance on C3 and C4 channels and the Beta frequency band, consistent with neurophysiological findings. This work advances EEG signal processing techniques and provides new insights and tools for BCI and neuroscience research.