Reproducibility and Replicability in Neuroimaging: Constrained IVA as an Effective Assessment Tool

Matrix decomposition techniques have been successfully applied in the analysis of multi-subject functional magnetic resonance imaging (fMRI) data. These data-driven approaches that assume the linear blind source separation (BSS) problem can yield an unsupervised and fully interpretable solution when there is a good model match. However, selecting a suitable model order that provides an accurate model match is an important challenge. Replicability and computational reproducibility are two key aspects that are also intimately related to interpretability. Despite clear evidence that solutions with poor reproducibility can lead to suboptimal results, the evaluation of reproducibility in matrix decomposition techniques remains limited in the existing literature. We propose the use of constrained independent vector analysis (cIVA), a state-of-the-art joint BSS technique, to assess the influence of model order selection for replicability and reproducibility. We demonstrate the attractiveness of clVA for replicability by alleviating permutation ambiguity as well as enabling additional quantification opportunities. Our results show that highly reproducible model orders achieve a good model match with highly interpretable and replicable solutions when clVA is applied to four different restina-state fMRI datasets.