A Multibranch of Convolutional Neural Network Models for Electroencephalogram-Based Motor Imagery Classification

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning enables automatic high-level feature extraction, optimizing efficiency in various applications.
• Convolutional Neural Network (CNN)-based electroencephalography (EEG) motor imagery classification has shown high accuracy but is limited by single convolution scales.
• The optimal convolution scale for EEG motor imagery classification varies significantly between subjects, impacting overall classification precision.

Purpose of the Study:

• To propose and evaluate multi-branch CNN models for improved EEG motor imagery classification.
• To address the limitation of single convolution scales in existing CNN models for EEG data.
• To enhance the extraction of spatial and temporal features from raw EEG data by utilizing multiple filter kernel sizes.

Main Methods:

• Development of multi-branch CNN architectures, including MBEEGNet and MBShallowConvNet.
• Utilizing different filter kernel sizes across parallel branches within the CNN models.
• Training and testing the proposed models on public datasets: BCI Competition IV 2a and High Gamma Dataset (HGD).

Main Results:

• The multibranch EEGNet (MBEEGNet) achieved a 9.61% improvement in classification accuracy compared to a fixed one-branch EEGNet.
• MBEEGNet also showed a 2.95% accuracy improvement over a variable EEGNet model.
• The multibranch ShallowConvNet (MBShallowConvNet) demonstrated a 6.84% accuracy increase over its single-scale counterpart.
• The proposed multi-branch models outperformed existing state-of-the-art methods in EEG motor imagery classification.

Conclusions:

• Multi-branch CNN models effectively extract spatial and temporal features from raw EEG data.
• The proposed approach overcomes the limitations of single convolution scales, leading to improved subject-specific classification accuracy.
• These advanced CNN models represent a significant advancement in EEG motor imagery classification technology.