Estimating Causality Measures from Reconstructed Source Time Courses When Large Background Activities Exist

Abstract

This paper proposes to use state-space modeling to incorporate the observational noise in the MVAR formulation. The difference of the proposed method from recently-reported investigations in the same direction is that the proposed method allows the noise to have a temporal structure, and thus the method can handle the background interference (brain noise) as the observational noise. The proposed method first estimates the MVAR coefficients of the brain noise from the control period, and then constructs the state-space model using these coefficients. The expectation maximization (EM) algorithm estimates the MVAR coefficients of the brain signals. The estimated MVAR coefficients are then used to compute Granger-causality-based measures such as the partial directed coherence and the directed transfer function. The results of numerical experiments that validate the method’s effectiveness are presented.