Existing spatial independent component analysis (ICA) methods for multi-subject fMRI datasets have mainly focused on detecting common components across subjects, under the assumption that all the subjects in a group share the same (identical) components. However, as a data-driven approach, ICA could potentially serve as an exploratory tool at multi-subject level, and help us uncover inter-subject differences in patterns of connectivity (e.g., find subtypes in patient populations). In this work, we propose a methodology named gRAICAR that exploits the data-driven nature of ICA to allow discovery of sub-groupings of subjects based on reproducibility of their ICA components. This technique allows us not only to find highly reproducible common components across subjects but also to explore (without a priori subject groupings) components that could classify all subjects into sub-groups. gRAICAR generalizes the reproducibility framework previously developed for single subjects (Ranking and averaging independent component analysis by reproducibility-RAICAR-Yang et al., Hum Brain Mapp, 2008) to multiple-subject analysis. For each group-level component, gRAICAR generates its reproducibility matrix and further computes two metrics, inter-subject consistency and intra-subject reliability, to characterize inter-subject variability and reflect contributions from individual subjects. Nonparametric tests are employed to examine the significance of both the inter-subject consistency and the separation of subject groups reflected in the component. Our validations based on simulated and experimental resting-state fMRI datasets demonstrated the advantage of gRAICAR in extracting features reflecting potential subject groupings. It may facilitate discovery of the underlying brain functional networks with substantial potential to inform our understandings of development, neurodegenerative conditions, and psychiatric disorders.
Copyright © 2012 Elsevier Inc. All rights reserved.