"Pre-Train, Prompt" Framework to Boost Graph Neural Networks Performance in EEG Analysis.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is crucial for neuroscience and clinical diagnosis but presents challenges due to complex non-Euclidean data structures and scarcity.
• Training effective graph neural network (GNN) models on limited EEG data is difficult, hindering advancements in EEG analysis.
• Existing methods often struggle with generalization and robustness when faced with limited EEG datasets.

Purpose of the Study:

• To develop a novel GNN framework, GNN-based EEG Prompt Learning (GEPL), to address data scarcity in EEG analysis.
• To enhance the performance of GNN models for EEG classification using limited target domain data.
• To improve the generalization, robustness, and interpretability of EEG analysis models.

Main Methods:

• Proposed a 'pre-train, prompt' framework (GEPL) for GNN-based EEG analysis.
• Employed unsupervised contrastive learning for pre-training on a large-scale EEG dataset.
• Utilized graph prompt learning to transfer knowledge to target EEG datasets with limited data.

Main Results:

• GEPL significantly outperformed traditional fine-tuning methods in classification accuracy and AUC across five EEG datasets.
• Demonstrated improved generalization, robustness, and computational efficiency compared to existing approaches.
• Significantly reduced overfitting risks associated with limited EEG data and provided interpretable results by highlighting relevant brain regions.

Conclusions:

• The 'pre-train, prompt' paradigm is highly effective for EEG analysis, particularly when data is limited.
• GEPL offers a promising solution for enhancing GNN performance in data-scarce domains.
• The framework's interpretability and efficiency suggest broad applicability in neuroscience and beyond.