Tensor Coreset Decomposition for Robust Dynamic Functional Connectivity Analysis

Dynamic functional network connectivity (dFNC) captures time-varying interactions among large-scale brain networks, revealing flexible coordination not apparent in static analyses. Studying these dynamics is important because fluctuations in connectivity reflect modular reorganization of networks and have been linked to psychiatric and neurological disorders. A common dFNC workflow segments fMRI data into windows and clusters the connectivity matrices into recurring brain states, but a key step is selecting representative windows. The popular exemplar pipeline uses variance–extrema selection (VES), which can overemphasize transients and yield instability. In this study, we propose a tensor coreset–based selection (TCS) scheme, based on tensor coreset decomposition, that deterministically chooses a compact, globally representative subset of windows by minimizing a kernel-based reconstruction error. Representative windows serve as centroids for clustering all windows into connectivity states. We applied TCS and VES to 600 resting-state fMRI scans from controls and schizophrenia patients. Both methods recovered similar state topologies, but TCS achieved higher clustering stability and more robust detection of group-level transition effects, with shorter windows. Thus, TCS is a practical and reliable scheme for exemplar selection in dFNC pipelines.