Discriminative Structured Feature Engineering for Macroscale Brain Connectomes.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Graph Theory

Background:

• Neuroimaging provides detailed in vivo measurements of brain connectivity, represented as high-dimensional matrices.
• Comparing these matrices between patient and control cohorts requires robust computational network models for feature engineering and statistical significance estimation.

Purpose of the Study:

• To develop a novel method for revealing intrinsic features of functional brain connectivity matrices with discriminative power for group comparisons.
• To enhance feature selection by preserving discriminative edges and reducing false positives through an optimization procedure.

Main Methods:

• The proposed method encourages co-selection of edges connected to the same node to maximize the preservation of discriminative edges.
• An optimization procedure evaluates the statistical significance of extracted edges to remove trivial ones, reducing the false positive rate.

Main Results:

• The novel approach outperformed ℓ1 regularized logistic regression, univariate t-test, and stability selection in feature selection and group comparison.
• The method increased the F-measure of feature selection and identified endogenous, discriminative connectivity patterns consistent with biomedical literature.

Conclusions:

• The developed data-driven method offers a new avenue for analyzing functional brain connectomes.
• It effectively identifies discriminative connectivity patterns, paving the way for improved understanding of brain network models in clinical populations.