Adaptive node feature extraction in graph-based neural networks for brain diseases diagnosis using self-supervised...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is crucial for diagnosing brain diseases.
• Brain networks derived from EEG signals offer valuable diagnostic insights.
• Current methods for defining EEG brain network node characteristics are inconsistent and inadequate.

Purpose of the Study:

• To develop a novel, adaptive method for extracting robust node features from EEG signals.
• To integrate these features into Graph Neural Networks (GNNs) for enhanced brain disorder prediction.
• To demonstrate the method's superiority and generality across various neurological conditions.

Main Methods:

• Utilized a task-induced self-supervised learning technique for adaptive node feature extraction from EEG.
• Combined novel node features with Pearson correlation-based edge features.
• Integrated the feature extraction module into common GNN architectures (GCN, GraphSAGE, GAT).

Main Results:

• The proposed method consistently outperformed existing feature selection techniques in brain disorder prediction tasks.
• Demonstrated high generality across diverse conditions including depression, schizophrenia, and Parkinson's disease.
• Revealed significant spatial patterns and identified pivotal brain regions through graph pooling and structural learning.

Conclusions:

• The novel adaptive method provides a superior and generalizable approach for EEG-based brain network analysis.
• This plug-in feature extraction module enhances the performance of GNNs for neurological disorder diagnosis.
• The approach offers deeper insights into the neural underpinnings of brain disorders by uncovering spatial patterns.