LINEAR SHRINKAGE PRECISION MATRIX METHOD FOR IMPROVED FUNCTIONAL CONNECTIVITY ESTIMATION IN NEUROIMAGING

Abstract

Neurological disorders, arising from abnormalities in the nervous system, impose a major global health burden due to the lack of curative treatments and challenges in early diagnosis. Resting-state functional MRI (rs-fMRI) enables non-invasive mapping of brain activity, but conventional functional connectivity (FC) estimation often fails to capture complex network interactions. This study combines machine learning with shrinkage-based FC estimation to identify discriminative features that may serve as biomarkers across multiple disorders.

Five FC methods—Pearson’s, Spearman’s, Empirical Covariance, Ledoit–Wolf, and Oracle Approximating Shrinkage—were compared using classifiers including Logistic Regression, SVM, Random Forests, KNN, Naive Bayes, and CNNs. L1-regularized models guided feature selection.

Shrinkage-based estimators outperformed traditional methods, and a proposed Weighted Connectivity Matrix further improved accuracy, particularly with interpretable classifiers. Synthetic data confirmed robustness, and sensitivity analysis showed greatest influence from ROI count, followed by sample size. Results highlight the potential of shrinkage-based FC approaches for neuroimaging-based classification.