BrainGridNet: A two-branch depthwise CNN for decoding EEG-based multi-class motor imagery.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) are vital for restoring interaction for individuals with disabilities.
• Decoding multi-class motor imagery (MI) tasks requires accurate, stable, and lightweight neural networks.

Purpose of the Study:

• To introduce BrainGridNet, a novel convolutional neural network (CNN) framework for decoding five-class MI tasks using 3D electroencephalography (EEG) data.
• To evaluate BrainGridNet's performance in terms of accuracy, computational efficiency, and robustness.

Main Methods:

• Developed BrainGridNet, a CNN integrating two intersecting depthwise CNN branches.
• Utilized 3D electroencephalography (EEG) data for decoding a five-class MI task.
• Assessed performance in time and frequency domains, including robustness to signal loss.

Main Results:

• Achieved 80.26% accuracy and a kappa value of 0.753, surpassing state-of-the-art models.
• Demonstrated superior performance in the frequency domain and optimal computational efficiency.
• Maintained robust accuracy even with the loss of 16 electrode signals.

Conclusions:

• BrainGridNet offers strong feature extraction, high decoding accuracy, and steady efficacy.
• Its low computational cost makes it suitable for real-time BCI applications.
• The framework effectively identifies discriminative features, critical brain regions, and frequency bands for MI classification.