A Transformer-Based Approach Combining Deep Learning Network and Spatial-Temporal Information for Raw EEG...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) signals possess complex spatial-temporal dependencies crucial for accurate classification.
• Transformer models, known for attention mechanisms, have potential in analyzing long-sequence data like EEG.
• Limited research exists on Transformer-based models for motor-imagery EEG classification, especially with cross-individual validation.

Purpose of the Study:

• To design and evaluate Transformer-based models for motor-imagery EEG classification.
• To assess the performance of these models using cross-individual validation.
• To visualize attention mechanisms for understanding EEG patterns and model interpretability.

Main Methods:

• Developed Transformer-based deep learning models utilizing the spatial-temporal characteristics of EEG signals.
• Trained and validated models on the PhysioNet dataset for motor-imagery tasks.
• Incorporated positional embedding modules and analyzed attention weight visualizations.

Main Results:

• Achieved classification accuracies of 83.31% (2-class), 74.44% (3-class), and 64.22% (4-class) using 3s EEG data in cross-individual validation.
• Outperformed state-of-the-art models by up to 2.11%.
• Attention weight visualization revealed event-related desynchronization (ERD) patterns, consistent with spectral analysis of sensorimotor rhythms.

Conclusions:

• Transformer-based models offer a powerful and novel approach for motor-imagery EEG classification.
• Positional embeddings enhance classification performance.
• The developed methods provide valuable tools for EEG analysis and have significant implications for brain-computer interface (BCI) systems.