LOCUS: A REGULARIZED BLIND SOURCE SEPARATION METHOD WITH LOW-RANK STRUCTURE FOR INVESTIGATING BRAIN CONNECTIVITY.

Area of Science:

• Neuroscience
• Network Science
• Data Science

Background:

• Network-oriented research is growing across scientific fields.
• Brain network connectivity measures are crucial for understanding brain organization and serve as neural fingerprints.
• Analyzing high-dimensional connectivity matrices presents challenges like unknown latent sources and spurious findings.

Purpose of the Study:

• To propose a novel blind source separation method, LOCUS (low-rank structure and uniform sparsity), for network measures.
• To address challenges in analyzing brain connectivity matrices, including dimensionality and spurious findings.
• To develop a data-driven decomposition method for enhanced network analysis.

Main Methods:

• Developed LOCUS, a blind source separation method incorporating low-rank structure and uniform sparsity.
• Introduced a novel angle-based uniform sparsity regularization for improved performance.
• Implemented an efficient iterative node-rotation algorithm to solve the non-convex optimization problem.

Main Results:

• LOCUS demonstrates more efficient and accurate source separation for connectivity matrices compared to existing methods.
• The novel sparsity regularization outperforms existing controls for low-rank tensor methods.
• Simulations confirm the advantages of LOCUS.
• Application to the Philadelphia Neurodevelopmental Cohort revealed novel, biologically insightful connectivity traits.

Conclusions:

• LOCUS provides a powerful, data-driven approach for analyzing brain network connectivity.
• The method effectively handles high-dimensional data and identifies previously undiscovered connectivity patterns.
• LOCUS advances neuroimaging analysis by offering more accurate and insightful decomposition of brain networks.