Bi-hemisphere Interaction Convolutional Neural Network for Motor Imagery Classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Decoding electroencephalogram (EEG)-based motor imagery Brain-Computer Interfaces (MI-BCI) is challenging due to signal complexity and inter-subject variability.
• Convolutional Neural Networks (CNNs) have improved MI-BCI accuracy, but often neglect crucial inter-hemisphere functional connections.

Purpose of the Study:

• To develop a novel CNN architecture for subject-independent MI-BCI decoding.
• To integrate inter-hemisphere functional connections into CNN models for enhanced spatial information utilization.

Main Methods:

• A CNN architecture was designed to explicitly model inter-hemisphere connections.
• Channel Average Referencing was applied to one hemisphere, compared with the contralateral hemisphere.
• Cosine similarity identified correlated channels, which were combined with the original hemisphere for spatial filtering.

Main Results:

• The proposed method demonstrated superior performance on the Cho2017 and OpenBMI datasets.
• The CNN architecture effectively learned inter-hemispheric connections, improving decoding accuracy.
• The technique aligns better with actual brain functionality during motor imagery tasks.

Conclusions:

• The novel CNN architecture significantly enhances subject-independent MI-BCI decoding by incorporating inter-hemisphere brain dynamics.
• This approach offers a more biologically plausible and effective method for Brain-Computer Interface development.
• Future research can explore further integration of neurophysiological principles into BCI algorithms.