Bayesian thresholded modeling for integrating brain node and network predictors.

Area of Science:

• Neuroscience
• Biostatistics
• Medical Imaging

Background:

• Neuroscience research increasingly integrates diverse brain imaging data (structure, function, metabolism).
• Current approaches often analyze imaging traits separately at localized node-level or network-level metrics.
• A gap exists in integrating these hierarchies for a comprehensive understanding of neurobiological mechanisms.

Purpose of the Study:

• To propose a novel Bayesian regression model for integrating multi-level brain imaging data (node and network metrics).
• To develop a method for characterizing the interplay between different neuroimaging components.
• To identify and quantify neuromarkers and their predictive mechanisms for phenotypic outcomes.

Main Methods:

• Developed a Bayesian regression model accommodating vector-variate and matrix-variate predictors.
• Introduced a joint thresholded prior to capture coupling, grouping, and spatial contiguity of signal patterns.
• Utilized posterior inference to identify neuromarkers and assess predictive uncertainty.

Main Results:

• The proposed Bayesian model significantly outperforms alternative methods in out-of-sample prediction and feature selection via simulations.
• The model successfully identified and quantified node- and network-level neuromarkers.
• Applied to children's general mental abilities, the model established a predictive mechanism using task contrast traits and resting-state sub-networks.

Conclusions:

• The novel Bayesian regression model effectively integrates multi-level brain imaging data for enhanced neurobiological insight.
• This approach provides a powerful tool for identifying predictive neuromarkers and understanding their mechanisms.
• The findings have implications for studying cognitive abilities and other phenotypic outcomes.