A Lightweight Dual-Attention Neural Network for Robust and Efficient EEG Motor Imagery Decoding
Guangying Wang1, Xipeng Song1, Lin Jiang1
1The Clinical Hospital of Chengdu Brain Science Institute, MOE Key Lab for Neuroinformation, University of Electronic Science and Technology of China, Chengdu 611731, P. R. China.
International Journal of Neural Systems
|March 19, 2026
Summary
A new lightweight Dual-Attention-EEGNet (DA-EEGNet) model improves motor imagery-based brain-computer interfaces (MI-BCI) by effectively modeling spatial-temporal features with minimal parameters. This efficient model offers a strong balance between accuracy and parameter count for MI-BCI applications.
Area of Science:
- Neuroscience
- Biomedical Engineering
- Machine Learning
Background:
- Motor imagery-based brain-computer interfaces (MI-BCI) require sophisticated models for spatial-temporal feature extraction.
- Existing deep learning models often face challenges with parameter efficiency and effective feature modeling.
Purpose of the Study:
- To propose a lightweight model, Dual-Attention-EEGNet (DA-EEGNet), for enhanced MI-BCI performance.
- To improve spatial-temporal feature modeling while maintaining a compact model size.
Main Methods:
- Developed DA-EEGNet by extending the EEGNet backbone with channel and depth attention modules.
- Evaluated the model on two MI benchmark datasets using subject-dependent, subject-independent, and dataset-independent classification strategies.
- Conducted ablation studies and visualization analyses (attention heatmaps, topographies) to assess model contributions and interpretability.
Main Results:
- DA-EEGNet achieved high average classification accuracies ([Formula: see text] and [Formula: see text]) with only 3.97k trainable parameters.
- The model outperformed or matched existing deep learning approaches with significantly larger parameter counts.
- Ablation studies confirmed the effectiveness of both channel and depth attention modules.
Conclusions:
- DA-EEGNet offers a favorable parameter-accuracy trade-off for MI-BCI.
- The model effectively captures neurophysiologically meaningful spatial-temporal patterns.
- DA-EEGNet serves as an efficient and interpretable baseline for future MI-BCI research and applications.
