Multi-scale convolutional transformer network for motor imagery brain-computer interface.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interface (BCI) systems translate neural signals into commands for external devices.
• Convolutional Neural Networks (CNNs) are used for decoding motor imagery electroencephalography (MI-EEG) signals.
• Existing CNN methods struggle with individual EEG variability and limited receptive fields.

Purpose of the Study:

• To introduce the Multi-Scale Convolutional Transformer (MSCFormer) model for improved EEG signal decoding in BCIs.
• To address the limitations of individual variability and feature extraction in traditional CNNs.
• To enhance the accuracy and robustness of EEG-based BCIs.

Main Methods:

• Developed MSCFormer, integrating multi-scale CNN branches for feature extraction and a Transformer module for global dependency capture.
• Utilized multi-branch CNNs to mitigate individual EEG signal variability and improve generalization.
• Employed a Transformer encoder for enhanced global feature integration and decoding performance.

Main Results:

• MSCFormer achieved 82.95% accuracy on the BCI IV-2a dataset and 88.00% on the BCI IV-2b dataset.
• Five-fold cross-validation yielded kappa values of 0.7726 (BCI IV-2a) and 0.7599 (BCI IV-2b).
• The model surpassed several state-of-the-art methods in decoding accuracy and robustness.

Conclusions:

• MSCFormer demonstrates superior performance in EEG-based BCI applications.
• The model's architecture effectively handles individual EEG variability and integrates multi-scale features.
• MSCFormer shows significant potential for advancing BCI technology and applications.