Spatio-Temporal Hypergraph Attention Networks for Brain Disease Analysis.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Functional brain connectivity networks are crucial for diagnosing neurological disorders.
• Existing vector or graph methods struggle to capture complex spatio-temporal network architectures.
• Current approaches lack priors for cross-window network interactions.

Purpose of the Study:

• To propose a novel spatio-temporal hypergraph attention network framework for brain network analysis.
• To enhance the characterization of intricate spatio-temporal topological architectures.
• To improve diagnostic performance for neurological disorders.

Main Methods:

• Developed a temporal attention network with temporal similarity priors for fMRI data.
• Designed a hierarchical hypergraph generation module for multi-scale modeling.
• Employed a spatial attention network with hypergraph message passing for spatial interactions.
• Utilized a multi-layer perceptron for classification.

Main Results:

• The proposed method effectively extracts long-range dependency information from fMRI.
• Achieved multi-scale modeling of high-order spatio-temporal structures.
• Demonstrated superior diagnostic performance on ADNI and PD datasets compared to state-of-the-art methods.
• Provided discriminative graph features for brain disease diagnosis.

Conclusions:

• The spatio-temporal hypergraph attention network framework offers a powerful new approach for brain network analysis.
• The method significantly improves diagnostic accuracy for neurological disorders.
• The framework effectively models complex spatio-temporal brain network dynamics.