Numerical Variability of functional MRI Graph Measures

Network neuroscience provides a powerful framework for studying the mechanisms underlying brain-related diseases. As analyses become increasingly computational, ensuring their numerical reliability has become a critical challenge. Small perturbations introduced during processing can propagate through complex pipelines, leading to variability in outcomes and raising concerns about the reproducibility of reported findings. Addressing this issue requires systematic evaluation of pipeline stability to ensure results remain within acceptable numerical limits. While the numerical variability of structural imaging workflows has been investigated, with findings ranging from negligible to substantial, functional MRI (fMRI) pipelines and their derived graph measures remain underexplored. Without rigorous stability assessment, conclusions drawn from these measures may remain uncertain. We systematically evaluated the numerical variability of graph measures of functional connectivity derived from the widely-used fMRIPrep pipeline and compared it to population variability. The resulting Numerical-Population Variability Ratio (NPVR) values typically ranged from 0.1 to 0.2 for most graph metrics, indicating a measurable influence of numerical variability on network-derived outcomes. NPVR values varied across brain regions, thresholding choices, and confound regression strategies. These findings highlight numerical variability as an important factor in functional network studies, particularly when examining subtle effects or working with small sample sizes.