A Novel Robust Sparse Granger Causality Inference Method and Its Application in MI EEG

In the cognitive neuroscience research, Granger causality-based directed brain network estimation has emerged as a powerful tool for exploring the potential cognitive information flow and the underlying brain neural mechanisms. However, the traditional approaches are sensitive to outliers, which may result the spurious network connections and hinder the wide application in the neural mechanism explorations. In this work, inspired by the outlier noise immunity characteristics of the Student's t-distribution, we propose a robust causal network estimation method based on the hypothesis of dual Student's t-distribution (DStu-GCA). Specifically, we employ the Student's t-distribution to constrain the coefficients and residuals of the multivariate autoregressive (MVAR) model respectively, enabling robust outlier suppression while obtaining a sparse network structure. To evaluate the effectiveness of the proposed DStu-GCA, we have conducted both the simulation and real motor imagery (MI) electroencephalogram (EEG) application experiments. The experimental results have consistently indicated that the DStu-GCA can effectively suppress noise and realize an efficient estimation of directed EEG brain network, which may provide a promising tool for the cognitive neuroscience research.