Self-attention-based convolutional neural network and time-frequency common spatial pattern for enhanced motor...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) based brain-computer interfaces (BCIs) show potential for neuro-rehabilitation.
• Decoding MI EEG signals is challenging due to individual brain variations, requiring better classification performance.

Purpose of the Study:

• To propose a self-attention-based Convolutional Neural Network (CNN) combined with time-frequency common spatial pattern (TFCSP) for enhanced MI classification.
• To address limited training data using a data augmentation strategy for MI EEG datasets.

Main Methods:

• A self-attention-based CNN extracts temporal and spatial EEG information, with attention modules calculating channel weights.
• Time-frequency common spatial pattern (TFCSP) extracts multiscale time-frequency-space features.
• Features from TFCSP and self-attention CNN are concatenated for final MI classification.

Main Results:

• The proposed method achieved mean accuracies of 79.28% on the BCI Competition IV IIa dataset.
• The method yielded a mean accuracy of 86.39% on the BCI Competition III IIIa dataset.
• Superior classification accuracy was observed compared to state-of-the-art methods.

Conclusions:

• The combination of self-attention CNN and TFCSP effectively utilizes time-frequency-space EEG information.
• This approach significantly enhances MI classification performance.
• The proposed method offers improved accuracy for practical neuro-rehabilitation applications.