Multiscale Spatial-Temporal Feature Fusion Neural Network for Motor Imagery Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery is a key paradigm in brain-computer interfaces (BCIs), crucial for assistive technologies.
• Current deep learning decoding methods for motor imagery signals are limited by single-scale convolutions, hindering information extraction.
• Efficient and stable BCI interaction relies heavily on precise decoding of motor imagery.

Purpose of the Study:

• To propose a novel end-to-end convolutional neural network, MSTFNet, for enhanced EEG classification of motor imagery.
• To address the limitations of single-scale convolutions in existing deep learning models.
• To improve the accuracy and robustness of decoding motor imagery signals for BCI applications.

Main Methods:

• Developed MSTFNet, a convolutional neural network featuring multiscale spatial-temporal feature fusion.
• MSTFNet incorporates feature enhancement, multiscale temporal feature extraction, spatial feature extraction, and a feature fusion module.
• Implemented a data augmentation strategy to enhance model performance and conducted cross-session and leave-one-subject-out experiments for validation.

Main Results:

• MSTFNet achieved 83.62% and 89.26% accuracy on BCI Competition IV 2a and 2b datasets, respectively.
• Achieved 86.68% accuracy on a laboratory dataset and 66.31% in leave-one-subject-out experiments on BCI Competition IV 2a.
• Demonstrated superior performance compared to several state-of-the-art methods in decoding EEG signals for motor imagery.

Conclusions:

• MSTFNet effectively decodes electroencephalography (EEG) signals for motor imagery tasks.
• The multiscale spatial-temporal feature fusion approach enhances information extraction from motor imagery signals.
• The proposed model shows robust capability and potential for advanced BCI applications.