Bayesian Estimation of Conditional Independence Graphs Improves Functional Connectivity Estimates.

Area of Science:

• Neuroscience
• Computational Biology
• Statistical Modeling

Background:

• Functional connectivity in the brain, often studied with fMRI, measures correlated neuronal activity.
• Covariance measures coupled activity but cannot distinguish direct from indirect effects.
• Partial correlation identifies direct connections by regressing out confounding variables.

Purpose of the Study:

• To propose a probabilistic generative model for estimating functional connectivity.
• To represent connectivity using both partial correlations and conditional independence graphs.
• To improve upon existing methods like graphical LASSO for partial correlation estimation.

Main Methods:

• Developed a probabilistic generative model for functional connectivity analysis.
• Estimated partial correlations and conditional independence graphs.
• Applied the model to resting-state fMRI data from twenty subjects.
• Utilized a Bayesian formulation for quantifying uncertainty in connectivity estimates.

Main Results:

• The proposed model outperforms graphical LASSO in estimating partial correlations.
• The model provides a richer representation of functional connectivity than partial correlations alone.
• Empirical results reveal a clear backbone of brain connectivity with quantifiable uncertainty.
• The Bayesian approach highlights limitations of deterministic methods in interpreting noisy fMRI data.

Conclusions:

• The probabilistic model offers a more comprehensive approach to functional connectivity analysis.
• Quantifying uncertainty is crucial for accurate interpretation of brain connectivity from fMRI data.
• This method advances the understanding of direct and indirect neural interactions and their reliability.