MI-CAT: A transformer-based domain adaptation network for motor imagery classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is crucial for motor imagery (MI) brain-computer interfaces (BCIs) due to its safety and convenience.
• Deep learning, particularly Transformer models, shows promise for EEG signal decoding owing to their ability to focus on global information.
• Subject-to-subject variability in EEG signals presents a significant challenge for cross-domain generalization in BCIs.

Purpose of the Study:

• To address the challenge of subject-specific EEG signal variations in Transformer-based BCIs.
• To propose a novel architecture, MI-CAT, for effective cross-domain knowledge transfer in MI-BCIs.
• To enhance the classification performance of single-subject EEG decoding by leveraging data from multiple subjects.

Main Methods:

• Developed MI-CAT, a novel architecture utilizing Transformer's self-attention and cross-attention mechanisms.
• Employed a patch embedding layer to segment source and target EEG features into patches.
• Integrated multiple Cross-Transformer Blocks (CTBs) for adaptive bidirectional knowledge transfer and domain-specific attention blocks for feature alignment.

Main Results:

• Achieved competitive classification accuracy on public datasets: 85.26% on Dataset IIb and 76.81% on Dataset IIa.
• Demonstrated effective feature interaction and alignment between source and target domains using cross-attention.
• Validated the model's capability in capturing domain-dependent information for improved EEG decoding.

Conclusions:

• MI-CAT offers a powerful solution for decoding EEG signals in motor imagery BCIs.
• The proposed architecture effectively mitigates the impact of subject-specific EEG signal differences.
• This work facilitates the advancement of Transformer applications in brain-computer interface development.