|
Abstract:
|
The sliding-window approach to dynamic brain functional connectivity is limited by the choice of window length and relies on ad-hoc clustering to identify recurring connectivity states. An alternative based on Markov-switching models can estimate both dynamic states and connectivity graphs simultaneously, and is adaptive to changes at different time scales. However, classical estimation of these models assumes a fixed, prespecified number of states which is typically unknown a priori. We propose a nonparametric Bayesian approach to inferring dynamic directed connectivity in brain signals, using a switching vector autoregressive (SVAR) model with hierarchical Dirichlet process (HDP) prior on the evolution of states. By applying a HDP prior on the state transition probability measures over an infinite parameter space, the HDP-SVAR allows us to learn adaptively an unknown number of latent states in the time-evolving connectivity structure. An automatic relevance determination (ARD) prior on the VAR parameters is used to induce sparse connectivity graphs. Posterior inference is obtained by Gibbs sampling. The method is applied to EEG data to detect dynamic brain states evoked by tasks.
|
