Hierarchical Region-Network Sparsity for High-Dimensional Inference in Brain Imaging

Area of Science:

• Neuroimaging
• Statistical modeling
• Machine learning

Background:

• Structured sparsity penalization enhances statistical models for high-dimensional data.
• Prior work has studied brain architecture at the level of functional segregation (regions) and functional integration (networks) separately.

Purpose of the Study:

• To incorporate hierarchical priors on brain region-network architecture into logistic regression models.
• To improve the classification and recovery of neural activity effects in medical imaging.
• To bridge the gap between studying brain regions and brain networks.

Main Methods:

• Developed hierarchical region-network priors for logistic regression.
• Applied these priors to analyze neural activity patterns associated with 18 psychological tasks.
• Compared the performance of hierarchical priors against other sparse estimators.

Main Results:

• Hierarchical region-network priors demonstrated superior performance in classifying and recovering psychological tasks compared to other sparse estimators.
• Varying the emphasis on region versus network structure within the penalty captured distinct aspects of neural activity.
• The approach showed advantages in generalization performance, sample complexity, and domain interpretability.

Conclusions:

• Hierarchical priors integrating region and network information improve brain activity analysis.
• Neurobiological knowledge at both local and global levels enhances model performance.
• This method offers a more interpretable and efficient way to analyze complex brain data.