Identifying Network Correlates of Brain States Using Tensor Decompositions of Whole-Brain Dynamic Functional Connectivity

Network organization is fundamental to the human brain and alterations of this organization by brain states and neurological diseases is an active field of research. Many studies investigate functional networks by considering temporal correlations between the fMRI signal of distinct brain regions over long periods of time. Here, we propose to use the higher-order singular value decomposition (HOSVD), a tensor decomposition, to extract whole-brain network signatures from group-level dynamic functional connectivity data. HOSVD is a data-driven multivariate method that fits the data to a 3-way model, i.e., connectivity x time x subjects. We apply the proposed method to fMRI data with alternating epochs of resting and watching of movie excerpts, where we captured dynamic functional connectivity by sliding window correlations. By regressing the connectivity maps' time courses with the experimental paradigm, we find a characteristic connectivity pattern for the difference between the brain states. Using leave-one-subject-out cross-validation, we then show that the combination of connectivity patterns generalizes to unseen subjects as it predicts the paradigm. The proposed technique can be used as feature extraction for connectivity-based decoding and holds promise for the study of dynamic brain networks.