Feature Whitening

Whitening matrices were fit only on training folds and applied to test folds to avoid leakage. The approach mapped coefficients learned in whitened space back to original ROI feature space. The method used ZCA-cor whitening matrices computed from standardized two-feature data matrices. The article proposes whitening anatomically meaningful pairs of correlated neuroimaging features to separate overlapping information while retaining predictive performance. Regularized ZCA-cor whitening allowed partial decorrelation when original correlations might contain disease-relevant information. Whitening was applied to left-right homologous regions and GM-CSF volumes within the same brain region.