Impact of Connectivity Granularity: A Comparison of ROI and Network-Level Approaches for Early Schizophrenia Classification

Abstract

While schizophrenia diagnosis relies on clinical interviews, there is growing interest in neuroimaging-based computational tools to aid classification. In particular, resting-state fMRI-derived functional connectivity has been explored as a potential biomarker. Here, we compare two common approaches to computing functional connectivity-region of interest (ROI)-level and brain network-level, and evaluate their predictive power for classifying first-episode schizophrenia patients, in contrast to most prior work focusing on chronic patients. The two approaches also differ in how many regions or components they use. We therefore compared them across a wide range of granularities: a functional Craddock parcellation computed at increasing sizes, from K = 10 to K = 950 regions, alongside the AAL90 atlas and ICA with 9, 17 and 80 components. The impact of granularity outweighed that of the parcellation method. Accuracy rose with granularity and plateaued at roughly 70 regions, with no further gain or loss at finer resolutions. Above the plateau the method mattered little: AAL90 was strongest (72.4% across classifiers, 79.6% for its best model) and an 80-component ICA matched it. Overall, our findings highlight region-based connectivity from a medium-resolution atlas as a promising representation for early-stage schizophrenia classification. These findings replicated when the trained classifiers were transferred to a dataset from another site.