A Multi-Branch 3D Convolutional Neural Network for EEG-Based Motor Imagery Classification.

Area of Science:

• Neuroscience and Biomedical Engineering
• Machine Learning for Brain-Computer Interfaces

Background:

• Motor imagery (MI) classification using electroencephalogram (EEG) data faces challenges in preserving both temporal and spatial features.
• Existing EEG representation methods often struggle to capture multi-dimensional information effectively.

Purpose of the Study:

• To develop a novel framework for motor imagery classification that enhances feature representation.
• To improve the accuracy, robustness, and practicality of EEG-based MI classification.

Main Methods:

• Introduction of a new 3D representation for EEG signals, transforming them into a sequence of 2D arrays.
• Development of a multi-branch 3D convolutional neural network (3D CNN) tailored for the 3D EEG representation.
• Implementation of a corresponding classification strategy designed for the proposed multi-branch 3D CNN.

Main Results:

• The proposed framework achieved state-of-the-art classification kappa values.
• Demonstrated a 50% decrease in the standard deviation across subjects, indicating excellent robustness.
• Showcased high performance even with a reduced number of nine sampling electrodes, enhancing practicality.

Conclusions:

• The novel 3D EEG representation and multi-branch 3D CNN framework significantly improve motor imagery classification.
• The framework offers superior performance, robustness, and practicality compared to existing algorithms.
• The multi-branch structure effectively mitigates overfitting and reduces latency, crucial for small EEG datasets.