Dynamic graph attention network based on multi-scale frequency domain features for motion imagery decoding in...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) are crucial for upper limb rehabilitation in hemiplegic patients.
• Individual variability in motor imagery electroencephalogram (MI-EEG) signals hinders BCI decoding performance.
• Existing MI-based BCI decoding methods struggle with generalization to new patients.

Purpose of the Study:

• To propose a novel Multi-scale Frequency domain Feature-based Dynamic graph Attention Network (MFF-DANet) for improved upper limb motor imagery decoding.
• To address the challenge of individual variability in MI-EEG signals for better BCI generalization.
• To enhance the accuracy and interpretability of MI-based BCIs in hemiplegic rehabilitation.

Main Methods:

• Developed MFF-DANet utilizing multi-scale convolutional kernels for frequency band feature extraction.
• Implemented channel attention-based average pooling to retain critical frequency domain features.
• Integrated a graph attention convolutional network with electrode position priors to capture spatial topological features.

Main Results:

• Achieved optimal decoding accuracies of 61.6% (within-subject) and 52.7% (cross-subject) on the PhysioNet dataset.
• t-SNE visualization confirmed the effectiveness of MFF-DANet's designed modules.
• Adjacency matrix visualization demonstrated the physiological interpretability of extracted spatial topological features.

Conclusions:

• MFF-DANet effectively decodes upper limb motor imagery signals in hemiplegic patients.
• The proposed method shows promise in overcoming the generalization issues of current BCI decoding techniques.
• The integration of multi-scale frequency features and graph attention networks offers a physiologically interpretable and accurate approach for BCI applications.