Reproducibility in Joint Blind Source Separation: Application to fMRI Analysis.

Joint blind source separation (JBSS) techniques have been successfully applied for the analysis of multi-subject functional magnetic resonance imaging (fMRI) data. However, convergence in JBSS can be only guaranteed to a local optimum, since typically cost functions are non-convex. Also, iterative methods are usually implemented with random initialization for best performance, resulting in high variability, especially for more flexible solutions. Yet, the assessment of the reproducibility of JBSS has been limited in the literature, even though it has been demonstrated that when not taken into account, the solutions can be highly suboptimal. In this work, we propose a framework for the evaluation of the reproducibility of independent vector analysis, an important JBSS solution. We introduce a mechanism for selecting the model complexity that offers the most consistent and accurate solution, and demonstrate results to underline its importance using resting state fMRI data.