AbstractRecent neuroscience studies suggest that flexible changes in functional brain networks are associated with cognitive functions. Therefore, the technique that detects changes in dynamical brain structures, which is called “dynamic functional connectivity (DFC) analysis”, has become important for the clarification of the crucial roles of functional brain networks. Conventional methods analyze DFC applying static indices based on the correlation between each pair of time-series data in the different brain areas to estimate network couplings. However, correlation-based indices lead to incorrect conclusions contaminated by spurious correlations between time-series data. These spurious correlation issues of network analysis could be reduced by performing the analysis assuming data structures based on a relevant model. Therefore, we propose a novel approach that combines the following two methods: (1) model-based network estimation assuming a dynamical system for time evolution, and (2) sequential estimation of model parameters based on Bayesian inference. We, thus, assumed that the model parameters reflect dynamical structures of functional brain networks. Moreover, by given the model parameter as prior distribution of the Bayesian inference, the network changes can be quantified based on the comparison between prior and posterior distributions of model parameters. In this comparison, we used the Kullback-Leibler (KL) divergence as an index for such changes. To validate our method, we applied it to numerical data and electroencephalographic (EEG) data. As a result, we confirmed that the KL divergence increased only when changes in dynamical structures occurred. Our proposed method successfully estimated both network couplings and change points of dynamic structures in the numerical and EEG data. The results suggest that our proposed method is useful in revealing the neural basis of dynamic functional networks.Author summaryWe proposed a method for detecting changes in dynamical brain networks. Although the detection of temporal changes in network dynamics from neural data has become more important (aiming to elucidate the role of neural dynamics in the brain), an adequate method for detecting the time-evolving dynamics of brain networks from neural data is yet to be established. To address this issue, we proposed a new approach to the detection of change points of dynamical network structures of the brain combining data-driven estimation of a coupled phase oscillator model and sequential Bayesian inference. As the advantage of applying Bayesian inference, by given the model parameter as the prior distribution, the extent of change can be quantified based on the comparison between prior and posterior distributions. Specifically, by using the Kullback-Leibler divergence as an index for change in the dynamical structures, we could successfully detect the neuroscientifically relevant dynamics reflected as changes from prior distribution of model parameters. The results indicate that the model-based approach for the detection of change points of functional brain networks would be convenient to interpret the dynamics of the brain.