Our work deals with effective connectivity to detect and quantify relations between cerebral structures involved in the initiation and the diffusion of epileptic seizures, aiming at establishing flow information propagation graphs. We study different approaches to answer two questions: (i) the identification of uni- and bi-directional relations, (ii) the discrimination between direct and indirect links. Firstly, we investigate the Granger causality index as well as its extended frequential and/or conditional versions, before exploiting a phase slope index and introducing a new indicator based on partial directed coherence. Then, we focus on transfer entropy selected as a nonlinear and nonparametric method computed from two signals. This method is considered in its conditional form to detect direct links taking into account the presence of a third signal. Since this technique is sensitive to calibration parameters such as the model order, a "greedy" strategy is proposed to optimize the order estimation based on the Bayesian information criterion. All approaches are evaluated and compared using Monte Carlo experiments on linear and nonlinear autoregressive models and also on physiology-based models and real signals recorded on an animal model (guinea-pig) during a particular phase of a seizure corresponding to a narrowband tonic activity. Results on simulated signals allow us to establish coherent and consistent propagation graphs. For the real signals, without any ground-truth, which makes the assessment difficult, the use of surrogate data allows us to speculate a good behavior of our techniques and, for the three approaches tested, results appear coherent.
