On the Deep Learning Models for EEG-Based Brain-Computer Interface Using Motor Imagery.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Motor imagery (MI) brain-computer interfaces (BCI) are crucial for assistive technologies.
• Deep learning models show promise for classifying MI electroencephalography (EEG) signals.
• Direct comparison of existing MI-BCI deep learning models is challenging due to varied datasets and availability.

Purpose of the Study:

• To systematically evaluate and compare the performance of five recent deep learning models for MI-EEG classification.
• To identify the most effective deep learning models for BCI applications.
• To provide a benchmark for future MI-BCI classifier development.

Main Methods:

• Selected five prominent deep learning models: EEGNet, Shallow & Deep ConvNet, MB3D, and ParaAtt.
• Tested these models on two large, publicly available MI-EEG datasets.
• Included datasets from 42 and 62 human subjects.
• Evaluated model performance based on classification accuracy and training cost.

Main Results:

• Deep learning models exhibited similar performance on one dataset.
• EEGNet achieved the best performance on the second dataset.
• EEGNet demonstrated this superior performance with relatively low training costs.

Conclusions:

• EEGNet is a highly effective deep learning model for motor imagery BCI classification.
• The findings offer valuable insights for selecting and designing high-performing BCI classifiers.
• Standardized evaluation on large, public datasets is essential for comparing BCI models.