Sparse Graphical Models for Functional Connectivity Networks: Best Methods and the Autocorrelation Issue.

Area of Science:

• Neuroimaging
• Network Neuroscience
• Computational Neuroscience

Background:

• Sparse graphical models are widely used for analyzing functional brain networks from neuroimaging data.
• The performance of these models in brain network analysis requires detailed investigation.

Purpose of the Study:

• To compare various sparse graphical model estimation procedures and selection criteria for brain networks.
• To assess the impact of autocorrelation and whitening on functional brain network estimation.
• To validate findings using resting-state functional magnetic resonance imaging (fMRI) data.

Main Methods:

• Simulation studies with varying dimensions, sample sizes, data types, and sparsity levels.
• Evaluation of estimation procedures like smoothly clipped absolute deviation (SCAD).
• Comparison of selection criteria including Bayesian information criterion (BIC) and cross-validation (CV).

Main Results:

• The SCAD estimation method combined with BIC and CV selection demonstrated superior performance in identifying true connections and minimizing false positives.
• Autocorrelation in the data negatively impacts network estimation accuracy.
• The CV selection method can effectively mitigate the adverse effects of autocorrelation.

Conclusions:

• The SCAD method with BIC/CV selection is recommended for robust functional brain network analysis.
• Findings challenge the validity of previous fMRI studies using less optimal graphical models.
• Proper handling of autocorrelation is crucial for accurate brain network estimation in fMRI studies.