Learning High-Order Relationships with Hypergraph Attention-based Spatio-Temporal Aggregation for Brain Disease...

Area of Science:

• Neuroscience
• Computational Biology
• Medical Imaging

Background:

• Functional magnetic resonance imaging (fMRI) typically models pairwise brain region interactions, limiting the characterization of complex, high-order relationships.
• Existing hypergraph methods often use fixed structures and ignore temporal dynamics, reducing their effectiveness and interpretability for brain network analysis.

Purpose of the Study:

• To develop a novel framework for jointly learning informative, sparse high-order brain structures and their temporal dynamics from fMRI data.
• To improve the modeling of complex brain interactions for enhanced disease classification and understanding of neuropsychiatric disorders.

Main Methods:

• A novel framework integrating hypergraph structure learning with temporal dynamics using an information bottleneck principle.
• Key components include a multi-hyperedge binary mask module, a hypergraph self-attention aggregation module, and a spatio-temporal network for feature extraction.

Main Results:

• The proposed method achieves competitive performance against state-of-the-art approaches on benchmark fMRI datasets.
• The framework effectively captures meaningful high-order brain interactions and their temporal dynamics.

Conclusions:

• The developed hypergraph-based framework offers a more expressive and interpretable approach to modeling brain networks compared to traditional methods.
• This approach shows significant potential for advancing the analysis of brain connectivity in neuropsychiatric disorders.