EISATC-Fusion: Inception Self-Attention Temporal Convolutional Network Fusion for Motor Imagery EEG Decoding
Summary
The EISATC-Fusion model enhances motor imagery brain-computer interface (MI-BCI) accuracy using electroencephalography (EEG) by integrating advanced deep learning techniques. This novel approach improves decoding performance for human-machine interaction.
Area of Science:
- Neuroscience
- Computer Science
- Biomedical Engineering
Background:
- Motor imagery brain-computer interfaces (MI-BCI) leverage electroencephalography (EEG) for human-machine interaction.
- Decoding accuracy in MI-BCI is often limited by EEG signal non-stationarity and inter-subject variability.
Purpose of the Study:
- To propose the EISATC-Fusion model for improved MI EEG decoding.
- To enhance the accuracy and robustness of MI-BCI systems.
Main Methods:
- The EISATC-Fusion model combines a DS Inception block for multi-scale frequency analysis, a cnnCosMSA module for attention mechanisms, and a depthwise separable convolution-enhanced Temporal Convolutional Network (TCN).
- Layer fusion (feature and decision fusion) and a two-stage training strategy with early stopping were employed.
- Model interpretability was assessed using weight visualization.
Main Results:
- Achieved within-subject accuracies of 84.57% (Dataset 2a) and 87.58% (Dataset 2b).
- Attained cross-subject accuracies of 67.42% (two sessions) and 71.23% (one session) on Dataset 2a via transfer learning.
- Demonstrated model interpretability through weight visualization.
Conclusions:
- The EISATC-Fusion model significantly improves MI EEG decoding accuracy and robustness.
- The proposed model offers a promising advancement for practical MI-BCI applications.
