DBConformer: Dual-Branch Convolutional Transformer for EEG Decoding.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Electroencephalography (EEG)-based brain-computer interfaces (BCIs) are crucial for communication, but current models struggle with long-range temporal dependencies and inter-channel relationships.
• Convolutional Neural Networks (CNNs) and existing CNN-Transformer hybrids have limitations in capturing global EEG signal characteristics.

Purpose of the Study:

• To introduce DBConformer, a dual-branch convolutional-Transformer network designed to overcome the limitations of existing models for EEG decoding.
• To improve the accuracy, robustness, and interpretability of EEG decoding by effectively modeling both temporal dynamics and spatial patterns.

Main Methods:

• Developed a dual-branch network integrating a temporal Conformer for long-range temporal dependencies and a spatial Conformer for inter-channel interactions.
• Incorporated a lightweight channel attention module to refine spatial representations by prioritizing informative EEG channels.
• Evaluated DBConformer on motor imagery, seizure detection, and steady-state visual evoked potential paradigms across four settings.

Main Results:

• DBConformer consistently outperformed 13 competitive baseline models across all evaluated settings and paradigms.
• Achieved superior performance with over an eight-fold reduction in parameters compared to high-capacity EEG Conformer architectures.
• Visualization confirmed that extracted features are physiologically interpretable and align with existing knowledge.

Conclusions:

• DBConformer offers a significant advancement in EEG decoding, providing accurate, robust, and explainable results.
• The model's efficiency and interpretability make it a reliable tool for various BCI applications.
• The proposed architecture effectively integrates local and global feature extraction for enhanced EEG signal analysis.