Independent component analysis (ICA) is a data-driven approach frequently used in neuroimaging to model functional brain networks. Despite ICA’s increasing popularity, methods for replicating published ICA components across independent datasets have been underemphasized. Traditionally, the task-dependent activation of a component is evaluated by first back-projecting the component to a functional MRI (fMRI) dataset, then performing general linear modeling (GLM) on the resulting timecourse. We propose the alternative approach of back-projecting the component directly to univariate GLM results. Using a sample of 37 participants performing the Multi-Source Interference Task, we demonstrate these two approaches to yield identical results. Furthermore, while replicating an ICA component requires back-projection of component beta-values (βs), components are typically depicted only by t-scores. We show that while back-projection of component βs and t-scores yielded highly correlated results (ρ=0.95), group-level statistics differed between the two methods. We conclude by stressing the importance of reporting ICA component βs so – rather than component t-scores – so that functional networks may be independently replicated across datasets.
展开▼
