GEFM: Graph-Enhanced EEG Foundation Model.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electroencephalography (EEG) signals are crucial for disease diagnosis and healthcare.
• Labeled EEG data scarcity is a major limitation for developing effective diagnostic tools.
• Foundation models pre-trained on large unlabeled datasets offer a solution for diverse EEG tasks.

Purpose of the Study:

• To develop a novel foundation model for EEG that integrates both temporal dynamics and inter-channel relationships.
• To address the limitations of existing EEG foundation models that primarily focus on temporal information.
• To enhance the performance of EEG analysis across various downstream tasks.

Main Methods:

• Proposed the Graph-Enhanced EEG Foundation Model (GEFM) architecture.
• Integrated Graph Neural Networks (GNNs) to capture relational structures between EEG channels.
• Employed a masked autoencoder for efficient pre-training on large-scale unlabeled EEG data.
• Evaluated GEFM using three downstream tasks and compared various GNN architectures, including GCN.

Main Results:

• The GEFM consistently outperformed baseline methods across all evaluated downstream tasks.
• The GCN architecture within GEFM, with optimized configurations, showed particularly strong performance.
• Integrating inter-channel relationships alongside temporal dynamics proved beneficial for EEG analysis.

Conclusions:

• The proposed GEFM serves as a robust foundation model for EEG analysis.
• Incorporating inter-channel relationships is vital for advancing EEG foundation models.
• GEFM offers a promising approach to overcome data scarcity and improve EEG-based healthcare applications.