Deep Learning of Motor Imagery EEG Classification for Brain-Computer Interface Illiterate Subject

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-Computer Interface (BCI) illiterate subjects struggle with traditional algorithms due to weak neural activity.
• Existing frequency band-based methods often fail to achieve high accuracy for these individuals.

Purpose of the Study:

• To design and evaluate a deep learning model, specifically a convolutional neural network (CNN), for classifying motor imagery EEG signals in BCI illiterate subjects.
• To improve classification accuracy beyond the 70% threshold for BCI illiterate users.

Main Methods:

• A CNN was designed to automatically extract relevant features from electroencephalography (EEG) data.
• The CNN model was used for end-to-end classification of motor imagery for BCI illiterate subjects.
• Performance was compared against the conventional Common Spatial Patterns (CSP) combined with Linear Discriminant Analysis (LDA) algorithm.

Main Results:

• The CNN approach demonstrated an average classification accuracy increase of 18.4% compared to CSP+LDA.
• Accuracies exceeding 70% were achieved by the CNN for 9 out of 11 BCI illiterate subjects.
• CNNs require minimal prior knowledge, enabling the extraction of features beyond traditional frequency bands.

Conclusions:

• Convolutional neural networks offer a promising solution for improving BCI performance in subjects with limited neural signal modulation.
• The CNN's ability to learn complex features without predefined constraints is key to its success.
• Future research will focus on visualizing CNN-extracted features to enhance model interpretability.