Exploration of the optimal group-discriminating features using CC-ICA

A coefficient-constrained independent component analysis (CC-ICA) framework for second-level group analysis is proposed, which incorporates group membership information as a constraint into the mixing coefficients. Applications to simulated signals and hybrid fMRI data show that, compared with regular ICA, CC-ICA improves both the decomposition accuracy and the extraction sensitivity to group differences. CC-ICA is then applied to real fMRI data to explore the optimal tasks and features from 15 task combinations. Results are consistent with and extend various neuroimaging studies and may prove especially important for the identification of relevant biomarkers of brain disorders.