Time-Distributed Attention Network for EEG-Based Motor Imagery Decoding From the Same Limb.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) offer intuitive control but classifying motor imagery (MI) from a single limb remains challenging.
• Existing methods struggle with the nuanced differentiation of multiple MI tasks within the same limb.

Purpose of the Study:

• To develop and validate a novel decoding method for classifying four distinct upper limb joint movements and a resting state using electroencephalography (EEG).
• To investigate the efficacy of a time-distributed attention network (TD-Atten) integrated with long short-term memory (LSTM) for MI decoding.

Main Methods:

• EEG data were recorded from 20 participants performing four upper limb motor imagery tasks and a resting state.
• A time-distributed attention network (TD-Atten) processed multiband Common Spatial Pattern (CSP) features, adaptively weighting classes and frequency bands.
• Long short-term memory (LSTM) and dense layers were employed for sequential information learning and classification.

Main Results:

• The proposed TD-Atten-LSTM method achieved 46.8% accuracy for 5-class (4 MI + rest) and 53.4% for 4-class (4 MI) scenarios.
• Performance surpassed existing baseline and deep learning approaches.
• Attention weight visualization highlighted the model's focus on alpha-band features, aligning with known brain activation patterns during MI.

Conclusions:

• The developed attention-based BCI decoding method demonstrates feasibility and interpretability for classifying fine-grained motor imagery tasks from a single limb.
• This approach shows significant potential for advancing intuitive control in neuroprosthetics and robotic arm applications.