Identifying dynamic effective connectivity states in fMRI based on time-varying vector autoregressive models

We propose a framework to estimate the transition of effective connectivity states in functional magnetic resonance imaging (fMRI), with the changing experimental conditions. The fMRI effective connectivity is traditionally assumed to be stationary across the entire scanning timecourse. However, recent evidence shows that it exhibits dynamic changes over time. In this study, we employ a non-stationary model based on time-varying autoregression (TV-VAR) to capture the dynamic effective connectivity, and K-means clustering to identify the change-points of the connectivity states. The TV-VAR parameters are estimated sequentially in time using the Kalman filtering and the expectation- maximization (EM) algorithm. The extracted directed connectivities between brain regions are then used as features to the K-means algorithm to be partitioned into a finite number of states and to produce the state change-points, assuming the task condition boundaries are unknown. Experimental results on motor-task fMRI data show the ability of the proposed method in estimating the state-related changes in the motor regions during the resting-state and active conditions, with low squared estimation errors. The estimated brain-state connectivity also reveals different patterns between the healthy subjects and the stroke patients.