Dynamic functional connectivity analysis using transition-point time-window and clustering with optimized number of clusters in fMRI study

Abstract

Elucidating the temporal dynamics of brain-states has become a key issue in fMRI analysis of functional connectivity in the brain. The sliding time-window and clustering approach is the most widely utilized method for analyzing dynamic alterations in connectivity. This approach represents the time-varying aspect of functional connectivity using windowed connectivity matrices and decodes representative brain-states using a clustering approach. Though very intuitive, results obtained using such an approach may be contaminated with spatiotemporal noises induced by arbitrarily defined parameters, such as size, offset of the sliding time-window, and the number of clusters. In this regard, these parameters should be determined by spatiotemporal aspects of the data themselves. Specifically, each windowed connectivity matrix should be constructed based on the duration of the specific brain-state, and brain-states should be decoded with the number that best describes the distribution of windowed connectivity matrices. In the present study, we propose a novel brain-state extraction algorithm based on the state transition (BEST) for analyzing dynamic connectivity in order to decode representative brain-states and their temporal variability with data-driven parameters. We used the spatial standard deviation of the brain map in order to detect brain-state transition time-points that enabled us to set time-windows on the duration of each specific brain-state. Furthermore, we utilized the Bayesian information criterion clustering method in order to estimate the number of brain-states. Such approaches enable the BEST to decode representative brain-states and their temporal pattern without any a priori knowledge.