Abstract:
|
In the dynamic functional brain networks, a popular assumption of the underlying mechanism is that the time-varying brain networks traverse a finite number of quasi-stable hidden brain states, which can be modeled by state-space methods. A shortcoming of the popular ML-based state-space models is that the total number of hidden states need to be prespecified. In most cases, it is determined ad hoc or based on some data-driven approaches. To better estimate the connectivity patterns, we want to model the number of hidden states and the transition between different states at the same time. In our study, we propose an improved estimation framework based on Ting et al. (2018). We use the Bayesian nonparametric inference of the switching vector autoregressive process (Fox et al., 2011) to model the dynamic connectivity patterns. It allows an unknown number of hidden states and also removes the parametric assumption of model formulations. We validated our method with simulated data and also applied it to the resting-state fMRI data from the human connectome project.
This is joint work with Hernando Ombao, Professor of Statistics at KAUST and Chee-Ming Ting, Visiting Scholar at KAUST.
|