Time-frequency-space transformer EEG decoding for spinal cord injury
Fangzhou Xu1, Ming Liu1, Xinyi Chen1
1International School for Optoelectronic Engineering, Qilu University of Technology (Shandong Academy of Sciences), Jinan, 250353 People's Republic of China.
Cognitive Neurodynamics
|December 23, 2024
Summary
This study introduces a novel time-frequency-spatial transformer for analyzing electroencephalographic (EEG) signals in spinal cord injury patients. The model achieves 93.56% accuracy in motor imagery classification, offering a promising tool for brain activity analysis.
Area of Science:
- Neuroscience
- Artificial Intelligence
- Biomedical Engineering
Background:
- Transformer neural networks with self-attention mechanisms demonstrate effectiveness across various fields.
- Electroencephalographic (EEG) signal analysis is crucial for understanding brain activity and developing pattern recognition models.
- Motor imagery (MI) tasks in spinal cord injury (SCI) patients present unique challenges for EEG analysis.
Purpose of the Study:
- To explore a multi-channel deep feature decoding method using self-attention for EEG signal analysis.
- To construct an effective motor imagery classification model for SCI patients utilizing transformer neural networks.
- To investigate the utility of self-attention mechanisms in integrating inter-channel and intra-channel EEG features.
Main Methods:
- A time-frequency-spatial transformer algorithm was developed for analyzing MI-based EEG signals.
- The model integrates inter-channel and intra-channel features using a self-attention mechanism.
- EEG signals from SCI patients underwent time-frequency and spatial domain analysis before input into the transformer network.
Main Results:
- The proposed time-frequency-spatial transformer achieved a peak classification accuracy of 93.56% for MI tasks.
- Construction of an attention matrix brain network revealed similarities to brain networks derived from raw EEG signals.
- Self-attention coefficient brain networks demonstrated potential for illustrating correlated connections and sample differences.
Conclusions:
- The self-attention mechanism effectively integrates multi-domain EEG features for enhanced pattern recognition.
- The developed transformer network provides a discriminative approach for analyzing brain activity in clinical settings.
- Attention coefficient brain networks offer valuable insights into brain network connectivity and functional differences.
