Multiscale Convolutional Transformer for EEG Classification of Mental Imagery in Different Modalities

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Transformer models are increasingly used for electroencephalogram (EEG) analysis.
• Existing EEG transformer models often overlook spatial and spectral information, focusing primarily on temporal attention.
• Current imagery-based brain-computer interfaces (BCIs) struggle with classifying diverse mental imagery types.

Purpose of the Study:

• To develop a generalized model for decoding various mental imagery modalities.
• To investigate neural representations across motor, visual, and speech imagery tasks.
• To improve the performance and class limitations of current BCI systems.

Main Methods:

• Designed an experimental paradigm involving motor, visual, and speech imagery tasks.
• Employed EEG source localization to analyze brain network connectivity.
• Proposed a multiscale convolutional transformer utilizing multi-head attention across spatial, spectral, and temporal domains.

Main Results:

• Achieved mental imagery classification accuracies of 0.62, 0.70, and 0.72 on private, BCI Competition IV 2a, and ASU datasets, respectively.
• Demonstrated superior performance compared to conventional deep learning models.
• Successfully decoded neural representations from diverse mental imagery tasks.

Conclusions:

• The multiscale convolutional transformer offers a promising approach for advanced BCI systems.
• This model addresses limitations in classifying multiple imagery types and enhances overall BCI accuracy.
• The findings contribute to overcoming class limitations and improving classification performance in BCI research.