BrainTGL: A dynamic graph representation learning model for brain network analysis.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Understanding human brain functional mechanisms relies on modeling dynamic characteristics in functional brain networks (FBNs).
• Existing models often overlook complex spatial and temporal correlations within the brain.
• Accurate modeling is crucial for diagnosing brain disorders and identifying subtypes.

Purpose of the Study:

• To propose a novel temporal graph representation learning framework, BrainTGL, for enhanced analysis of functional brain networks.
• To address limitations in current models by incorporating complex spatial and temporal correlations.
• To improve the accuracy of brain disease diagnosis and subtype identification.

Main Methods:

• Developed BrainTGL, a framework featuring temporal graph pooling to refine network data and dual temporal graph learning to capture spatio-temporal features.
• Evaluated BrainTGL on four datasets (HCP, ABIDE, NMU-MDD, NMU-BD) for classification (disease diagnosis/gender) and clustering (subtype identification) tasks.
• Compared BrainTGL against state-of-the-art methods like GroupINN and ST-GCN.

Main Results:

• BrainTGL demonstrated significant improvements, particularly in Autism Spectrum Disorder (ASD) diagnosis, outperforming existing methods.
• Achieved an average accuracy increase of 4.2% over GroupINN and 8.6% over ST-GCN in ASD diagnosis.
• Showcased enhanced performance in both disease diagnosis and subtype identification across multiple disorders.

Conclusions:

• Learning spatial-temporal representations of brain networks effectively models dynamic characteristics in FBNs.
• BrainTGL offers superior performance for brain disorder diagnosis and subtype identification compared to current methods.
• The framework also provides improvements in computational efficiency and convergence speed, reducing training costs.