A robust multi-branch multi-attention-mechanism EEGNet for motor imagery BCI decoding

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor-Imagery-based Brain-Computer Interfaces (MI-BCI) offer significant potential for assisting individuals with motor impairments.
• Deep learning (DL) techniques enhance Electroencephalography (EEG) decoding through automatic feature extraction.
• Intersubject variability in EEG data presents a major challenge for DL-based decoding models, leading to performance inconsistencies.

Purpose of the Study:

• To develop a robust deep learning model for EEG decoding in MI-BCI applications.
• To address the challenge of intersubject variability in EEG data.
• To improve the accuracy and reliability of MI-BCI systems.

Main Methods:

• Proposed a novel multi-branch multi-attention mechanism EEGNet model (MBMANet).
• Employed a multi-branch structure for diverse feature extraction.
• Integrated different attention mechanisms within each branch for adaptive weight adjustment and multi-level feature fusion.

Main Results:

• Achieved a four-classification accuracy of 83.18% and a kappa of 0.776 on the BCI Competition IV-2a dataset.
• Outperformed eight other Convolutional Neural Network (CNN)-based decoding models.
• Demonstrated consistent and robust performance across all nine subjects, indicating reliability.

Conclusions:

• The combination of multi-branch and multi-attention mechanisms enables adaptive learning of EEG features, effectively managing data variability.
• The MBMANet model offers a feasible solution for enhancing DL-based MI-BCI systems.
• The proposed model leads to more accurate decoding of motor intentions, reduced training costs, and improved overall MI-BCI utility and robustness.