Dynamic weighted hypergraph convolutional network for brain functional connectome analysis.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Hypergraph structures capture complex relationships in brain functional connectomes (FC) beyond simple graphs.
• Existing hypergraph neural networks (HGNNs) often use static hypergraphs, limiting their ability to represent dynamic brain networks.

Purpose of the Study:

• To propose a dynamic weighted hypergraph convolutional network (dwHGCN) framework for enhanced brain FC analysis.
• To develop a model that incorporates learnable hyperedge weights for dynamic hypergraphs.

Main Methods:

• Generated hyperedges using sparse representation and calculated hyper similarity as node features.
• Implemented a neural network where hyperedge weights adaptively update during training.
• Utilized functional magnetic resonance imaging (fMRI) data from the Philadelphia Neurodevelopmental Cohort for classification tasks.

Main Results:

• The dwHGCN framework demonstrated superior performance compared to existing HGNNs on classification tasks.
• The adaptive weighting strategy effectively assigned higher weights to discriminative hyperedges.
• The model enhanced the interpretability of brain FC by identifying key interactions among regions of interest (ROIs).

Conclusions:

• The proposed dwHGCN framework offers a powerful tool for learning and interpreting dynamic brain functional connectomes.
• The model's ability to handle dynamic structures and learnable weights advances HGNN applications in neuroimaging.
• dwHGCN shows promise for broader applications in neuroimaging requiring robust representation learning and interpretation.