Using topological data analysis to compare inter-subject variability across resting state functional MRI brain representations

In neuroimaging, extensive post-processing of resting-state functional MRI (rfMRI) data is necessary for its application and investigation in relation to brain-behavior associations. Such post-processing is used to derive brain representations, lower dimensional feature sets used for brain-behavior association studies. A brain representation involves a choice of dimension reduction (a parcellation into regions or networks) and a choice of feature type, such as spatial topography, connectivity matrix, amplitude. However, widespread variability in rfMRI brain representations has hindered both reproducibility and knowledge accumulation across the field. Brain representation choice effects measurements of inter-subject variability, which muddies the comparison and integration of findings. We leveraged persistent homology on the subject-space topologies induced by 34 different brain representations to enable direct comparison of brain representations in the context of individual differences. Our findings reveal the importance of considering feature type when comparing results derived from different brain representations, suggesting best practices for assessing the replicability and generalizability of brain-behavior research in rfMRI data.