A generative model of whole-brain effective connectivity

Area of Science:

• Computational neuroimaging
• Systems neuroscience
• Network neuroscience

Background:

• Inferring effective (directed) connection strengths from fMRI data in whole-brain models is a significant challenge.
• Existing methods like regression dynamic causal modeling (rDCM) scale to large networks but face interpretability and parameter estimation issues.
• Large-scale networks with thousands of connections are difficult to interpret, and precise estimation of all parameters is often hindered by insufficient data points per parameter.

Purpose of the Study:

• To introduce sparsity constraints into the variational Bayesian framework of rDCM for task-based fMRI.
• To enable efficient effective connectivity analyses in whole-brain networks without requiring a priori connectivity structure assumptions.
• To address the challenges of interpretability and parameter estimation in large-scale network analyses.

Main Methods:

• Incorporation of sparsity constraints into the variational Bayesian framework of regression dynamic causal modeling (rDCM).
• Derivation of variational Bayesian update equations for sparse rDCM.
• Validation using both simulated and empirical functional magnetic resonance imaging (fMRI) data.

Main Results:

• Demonstrated feasibility of inferring effective connection strengths from fMRI data in networks exceeding 100 regions and 10,000 connections.
• Achieved whole-brain effective connectivity inference in single subjects with a run-time under one minute using parallelized code.
• Sparse rDCM successfully prunes fully connected networks during model inversion without prior structural assumptions.

Conclusions:

• Sparse rDCM provides a computationally efficient and scalable solution for whole-brain effective connectivity inference from fMRI data.
• The method overcomes limitations of previous approaches, enabling precise parameter estimation and improved interpretability in large-scale neural networks.
• Potential applications include advancing connectomics research and clinical neuromodeling, such as patient phenotyping based on whole-brain network structure.