GAH-TNet: A Graph Attention-Based Hierarchical Temporal Network for EEG Motor Imagery Decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) using motor imagery (MI) show promise for rehabilitation and communication.
• Electroencephalography (EEG) signals present challenges due to low signal-to-noise ratio, non-stationarity, and inter-subject variability, hindering accurate decoding.
• Existing methods struggle to simultaneously capture spatial, local, and global patterns in EEG signals.

Purpose of the Study:

• To develop an advanced deep learning framework for enhanced decoding of motor imagery EEG signals.
• To address the limitations of current methods in modeling complex spatio-temporal dynamics and channel interactions in EEG data.

Main Methods:

• Proposed the Graph Attention-based Hierarchical Temporal Network (GAH-TNet) integrating spatial graph attention and hierarchical temporal encoding.
• Introduced the Graph Attention Temporal Encoding block (GATE) for spatial dependencies and short-term dynamics.
• Developed the Hierarchical Attention-guided Deep Temporal Feature Encoding block (HADTE) for local and global feature extraction via two-stage attention and temporal convolution.

Main Results:

• The GAH-TNet achieved superior classification accuracy on two public MI-EEG datasets, reaching 86.84% on BCI IV 2a and 89.15% on BCI IV 2b.
• Ablation studies confirmed the effectiveness of the GATE and HADTE components.
• The model demonstrated robust generalization capabilities across different subjects.

Conclusions:

• The proposed GAH-TNet framework effectively models the spatio-temporal dynamics and topological structure of MI-EEG signals.
• This hierarchical and interpretable approach offers a novel method for improving EEG motor imagery decoding performance.