Related Experiment Video
Updated: May 20, 2026
08:45
Mapping Cortical Dynamics Using Simultaneous MEG/EEG and Anatomically-constrained Minimum-norm Estimates: an Auditory Attention Example
Published on: October 24, 2012
Robust decoding for MI-EEG: a hybrid transformer network using multi-perspective collaborative attention and dynamic
Mei Wang1, Zhibo Gong1, Yujie Li2
1College of Artificial Intelligence and Computer Science, Xi'an University of Science and Technology, No. 48 Shangu Avenue, Xi'an, 710600 Shaanxi China.
Cognitive Neurodynamics
|May 19, 2026
Summary
This study introduces HATNet, a novel deep learning model for brain-computer interfaces (BCI) that enhances electroencephalogram (EEG) signal decoding. HATNet effectively reduces noise and adapts to signal variations, improving motor imagery (MI) and motor execution (ME) task classification.
Area of Science:
- Neuroscience and Artificial Intelligence
- Brain-Computer Interface (BCI) research
- Signal Processing and Machine Learning
Background:
- Deep learning models are widely used for decoding electroencephalogram (EEG) signals in brain-computer interface (BCI) systems.
- Existing hybrid architectures struggle with multi-channel noise reduction and adapting to EEG signal's non-stationarity.
- Limitations include ineffective noise elimination and poor adaptability to data distribution drifts in EEG signals.
Purpose of the Study:
- To propose a novel end-to-end hybrid attention Transformer network (HATNet) for improved EEG classification.
- To address the challenges of noise suppression and non-stationarity in EEG signals for BCI applications.
- To enhance the decoding performance for both motor imagery (MI) and motor execution (ME) tasks.
Main Methods:
- Developed HATNet, integrating a convolutional neural network for feature extraction and a Collaborative Attention Mechanism for spatial noise suppression.
- Implemented a Dynamic Hyperbolic Tangent module within the Transformer to adapt to real-time EEG data distribution drifts.
- Utilized cross-layer residual fusion for integrating global and local spatio-temporal features, evaluated on MI and ME datasets.
Main Results:
- HATNet achieved state-of-the-art performance on three primary MI datasets (BCIC-IV-2a, BCIC-IV-2b, OpenBMI) and one ME dataset (HGD).
- Subject-dependent accuracies reached up to 81.25% (MI) and 96.20% (ME).
- Subject-independent accuracies reached up to 80.79% (MI) and 73.95% (ME), demonstrating robustness.
Conclusions:
- HATNet effectively suppresses spatial noise and adapts to EEG signal non-stationarity, outperforming existing models.
- The model shows significant superiority and robustness in decoding both motor imagery and motor execution tasks.
- HATNet offers a promising advancement for practical BCI applications requiring accurate EEG signal interpretation.