Multi-scale EEG feature decoding with Swin Transformers for subject independent motor imagery BCIs.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-Computer Interfaces (BCIs) face challenges with subject variability and non-stationary EEG signals, hindering model generalization.
• Differences in neural activity, electrode placement, and noise degrade BCI performance, requiring extensive recalibration for reliable cross-user operation.

Purpose of the Study:

• To develop a novel Compact Convolutional Swin Transformer (CCST) model for subject-independent BCIs.
• To enhance model generalization and reliability across diverse users without extensive recalibration.

Main Methods:

• Utilized hierarchical window-based self-attention combined with convolutional feature extraction in the CCST model.
• Employed multi-scale feature representation to capture local electrode interactions and global temporal dependencies.
• Evaluated CCST on BCI Competition IV (2a, 2b) and PhysioNet MI datasets using Leave-One-Subject-Out (LOSO) cross-validation.

Main Results:

• Achieved state-of-the-art classification accuracies: 68.27% (BCI Comp IV 2a), 76.61% (BCI Comp IV 2b), and 71.70% (PhysioNet MI).
• Demonstrated statistically significant performance improvements over benchmark models via Wilcoxon signed-rank test with Bonferroni correction.
• CCST showed reduced parameters and FLOPs compared to full self-attention models, enhancing efficiency for real-time applications.

Conclusions:

• CCST provides a scalable and efficient framework for adaptive, subject-independent BCIs.
• The model's enhanced generalization capabilities are crucial for real-world BCI deployment.
• CCST shows promise for applications in neurorehabilitation, assistive technology, and cognitive training.