Graph neural network based on brain inspired forward-forward mechanism for motor imagery classification in...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interface (BCI) systems rely on electroencephalogram (EEG) signals for motor imagery (MI-EEG) tasks.
• Conventional deep learning (DL) methods struggle to represent the topological relationships between EEG electrodes, limiting decoding accuracy.
• Understanding brain network dynamics and neural signal transmission is crucial for advancing BCI technology.

Purpose of the Study:

• To introduce a novel deep learning framework, F-FGCN, integrating Graph Neural Networks (GCN) and the forward-forward (F-F) mechanism.
• To enhance EEG signal decoding performance by incorporating functional topological relationships and signal propagation mechanisms.
• To improve the accuracy of motor imagery classification in BCI applications.

Main Methods:

• Developed a F-FGCN framework inspired by the brain's forward-forward neuronal mechanism.
• Represented multi-channel EEG data as a network based on Pearson correlation coefficients to capture inter-channel associations.
• Employed a pre-trained GCN followed by fine-tuning, with the F-F model for advanced feature extraction and classification.

Main Results:

• The F-FGCN framework was evaluated on a four-class categorization task using the PhysioNet dataset.
• Achieved high classification accuracies of 96.11% at the subject level and 82.37% at the group level.
• Demonstrated superior performance compared to classical and state-of-the-art models in EEG decoding.

Conclusions:

• The F-FGCN framework effectively decodes EEG signals by leveraging topological relationships and signal propagation.
• The proposed method significantly enhances the performance of downstream classifiers for BCI applications.
• F-FGCN shows strong potential for advancing the capabilities and applications of brain-computer interfaces.