Comparing the Effective Connectivity Graphs Estimated by Granger Causality Index with Transfer Entropy: A Case Study on Autism Spectrum Disorders

In recent years, increasing attention has been paid to the study of brain connectivity in order to detect brain abnormalities and to raise awareness of brain disorders like Autism spectrum disorder (ASD). In these studies, the brain connectivity network is estimated and its graph parameters are extracted to aid researchers in analyzing brain function and its disorders during various tasks. Selecting the suitable effective connectivity estimator which is able to estimate linear and nonlinear causal relationships is an important issue in accurate estimation of effective connectivity network and exploring its disorders. In this paper, we address this issue and also investigate the effect of choosing the effective connectivity estimator on detected abnormalities of effective connectivity graph of ASD subjects. Two well-known effective connectivity estimators are used: transfer entropy (TE) and granger causality index (GCI). We first simulate three different networks whose their causal connections have different linearity conditions and compare the sensitivity and specificity of TE and GCI in each case. It is shown that except in completely linear networks, TE generally outperforms GCI in terms of both sensitivity and specificity. In the next step, each of TE and GCI is applied to an EEG dataset recorded during a face processing task from two groups of healthy control (He) individuals and people with ASD. The networks estimated from the subjects of two groups are compared in terms of average degree, average path length and total clustering coefficient. It can be seen that just the average degree is significantly different (higher) in healthy subjects than in ASD patients by using both TE and GCI. So the results of both TE and GCI are in accordance with the underconnectivity theory of ASD.