Hypercomplex Graph Neural Network: Towards Deep Intersection of Multi-Modal Brain Networks.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Data Science

Background:

• Multi-modal neuroimaging studies reveal brain network organization's link to behavior.
• Graph Neural Networks (GNNs) are emerging tools for analyzing complex brain network data.
• Challenges exist in effectively integrating diverse neuroimaging modalities due to complex dependencies.

Purpose of the Study:

• To develop a novel method for analyzing multi-modal brain networks.
• To overcome limitations of existing GNNs in handling heterogeneous inter-modal dependencies.
• To enhance the characterization of interplay among anatomical, functional, and physiological brain alterations.

Main Methods:

• Proposed a HyperComplex Graph Neural Network (HC-GNN) modeling multi-modal networks as hypercomplex tensor graphs.
• Conceptualized HC-GNN as a dynamic spatial graph with an adjacency matrix representing inter-modal associations.
• Utilized hypercomplex operations for cross-embedding and cross-aggregation to deepen multi-modal representation coupling.

Main Results:

• HC-GNN demonstrated superior classification performance across three datasets.
• The method showed strong scalability to various types of neuroimaging modalities.
• Statistical analysis of saliency maps identified potential disease biomarkers.

Conclusions:

• HC-GNN offers a powerful paradigm for multi-modal brain network research.
• The approach effectively integrates diverse neuroimaging data for enhanced analysis.
• This work advances the understanding of brain network organization and its behavioral correlates.