An efficient algorithm for estimating brain covariance networks.

Area of Science:

• Neuroscience
• Network Science
• Statistical Modeling

Background:

• Covariance networks reveal brain inter-relationships but traditional methods (partial correlations, pairwise analyses) lack consistency and are sensitive to tuning parameters.
• Existing methods like graphical LASSO (gLASSO) and Pearson pairwise correlations (PPC) present a specificity-sensitivity trade-off, leading to discrepancies with misspecified parameters.

Purpose of the Study:

• Propose a consistent covariance network estimator, maximizing network likelihood (MNL), that is robust to tuning parameters.
• Validate the theoretical consistency and performance of the MNL algorithm via simulation and compare it against gLASSO and PPC.

Main Methods:

• Developed the maximizing network likelihood (MNL) algorithm for consistent covariance network estimation.
• Conducted theoretical validation and simulation studies comparing MNL with gLASSO and PPC across various tuning parameters and sample sizes.

Main Results:

• The MNL algorithm demonstrated high specificity (≥0.94) across all sample sizes, with increasing sensitivity as sample size grew.
• gLASSO and PPC showed a specificity-sensitivity trade-off, indicating sensitivity to parameter choices.
• MNL application to case study data revealed reduced connections in impaired groups and enhanced identification of between-lobe connectivity compared to gLASSO.

Conclusions:

• The MNL algorithm is an effective and consistent approach for constructing covariance brain networks.
• MNL offers improved accuracy and robustness, particularly beneficial for large sample sizes in brain-wide analyses.
• This method enhances understanding of brain network organizational features in both healthy and pathological conditions.