Connectome embedding in multidimensional graph spaces.

Area of Science:

• Neuroscience
• Graph Theory
• Network Science

Background:

• Brain connectomes exhibit complex topological organization.
• Graph theory provides quantitative methods to analyze network structures.
• Understanding network properties in higher dimensions is crucial for brain research.

Purpose of the Study:

• To investigate brain networks in high-dimensional spaces using graph theoretic nodal properties.
• To explore the utility of machine learning for classifying brain regions based on network features.
• To develop a new framework for quantifying network features in high-dimensional spaces.

Main Methods:

• Utilized graph theory to define brain networks in up to 10-dimensional spaces.
• Generated structural and functional connectomes from 100 healthy subjects (Human Connectome Project).
• Employed machine learning (nonlinear Gaussian kernels) to classify sensory and association brain areas.

Main Results:

• Nodal properties showed significant correlations at whole-brain and subnetwork levels.
• Machine learning achieved 80-86% accuracy in classifying brain regions in a 10D space.
• The largest accuracy gains were observed from 2D to 3D spaces, with nonlinear kernels outperforming linear ones.
• Quantified high multidimensional Euclidean distances in default mode, frontoparietal, and temporal networks.

Conclusions:

• A novel high-dimensional framework for brain network analysis was proposed.
• This framework effectively differentiates between sensory and association brain networks.
• The findings suggest new avenues for uncovering complex brain network properties.