CCPD under sparsity and low-rank constraints: multi-frequency dynamic functional network connectivity analysis in...

This study aims to jointly extract group-shared connectivity patterns and group-specific temporal and frequency information from multi-frequency dynamic functional network connectivity (dFNC) tensors of healthy controls (HCs) and schizophrenia patients (SZs) using a coupled canonical polyadic decomposition (CCPD) approach. Based on 145 subjects (71 SZs and 74 HCs) from the COBRE dataset, multi-frequency dFNC tensors were constructed via group independent component analysis and a filter-banked connectivity framework. A novel sparse and low-rank constrained CCPD (SLRCCPD) model was proposed to decompose the dFNC tensors, incorporating L1-norm regularization to enhance significant connections and nuclear norm-based low-rank approximation to improve clustering quality. The results revealed significant connectivity differences between SZs and HCs within auditory, somatomotor, cognitive control, visual, and cerebellar networks across five shared dynamic modules. Clustering of group-specific time-frequency weights showed that SZs had significantly higher fractional and dwell time in State 3 at both low- and high-frequency bands, along with fewer state transitions across all bands compared to HCs. The proposed SLRCCPD framework effectively captures abnormal multi-band dynamic functional connectivity in schizophrenia, providing a new computational tool and empirical pathway for investigating brain network dynamics and mechanistic studies of the disorder.