A New Constrained CCPD Approach Applied to Multi-frequency Dynamic Functional Network Connectivity Analysis

Dynamic functional network connectivity (dFNC) analysis is widely used to study brain disorders like schizophrenia, but most previous researches disregard the frequency profiles. This paper constructs two three-way multi-frequency dFNC tensors for healthy controls (HCs) and schizophrenia patients (SZs) using group independent component analysis and a filter-banked connectivity approach. Coupled canonical polyadic decomposition (CCPD) under sparsity and low-rank constraints is then proposed for separating these tensors into shared connectivity loading matrix, group-specific time and frequency weights, and group-specific subject intensities. Results of the Open Center of Biomedical Research Excellence (COBRE) schizophrenia dataset demonstrate the advantage of the proposed method and reveal significant independent component networks (ICNs) connectivity differences and altered temporal dynamics in SZs, particularly prolonged state 3 occupancy at low and high frequency bands and fewer state transitions. These findings provide new insights into functional dynamics of schizophrenia.