Estimating the interaction among neural networks is an interesting issue in neuroscience. Some methods have been proposed to estimate the coupling strength among neural networks; however, few estimations of the coupling direction (information flow) among neural networks have been attempted. It is known that Bayesian estimator is based on a priori knowledge and a probability of event occurrence. In this paper, a new method is proposed to estimate coupling directions among neural networks with conditional mutual information that is estimated by Bayesian estimation. First, this method is applied to analyze the simulated EEG series generated by a nonlinear lumped-parameter model. In comparison with the conditional mutual information with Shannon entropy, it is found that this method is more successful in estimating the coupling direction, and is insensitive to the length of EEG series. Therefore, this method is suitable to analyze a short time series in practice. Second, we demonstrate how this method can be applied to the analysis of human intracranial epileptic electroencephalogram (EEG) recordings, and to indicate the coupling directions among neural networks. Therefore, this method helps to elucidate the epileptic focus localization.phase synchronization, coupling direction, conditional mutual information, Bayesian estimation, epileptic EEG The synchronization analysis in the dynamical system has been a focus of attention in various disciplines [1] . Synchronization phenomena have been found in physical systems and biological systems. A typical example is the cardio respiratory interaction [2] . The synchronization between brain areas is associated with human motion [3] , sleep [4] , learning [5] and consciousness [6] . However, excessive synchronization is likely to result in epileptic seizures [7] . Therefore, it is necessary to develop a method to detect the synchronization occurrences, in particular the coupling direction between brain areas. Rosenblum and his colleagues [1,2] put forward two methods to identify the coupling direction: evolution map approach (EMA) and instantaneous period approach (IPA). The two methods detect a driving or casual relationship through the phase dynamics between two channel signals. EMA and IPA have been successfully utilized to detect the coupling direction between the heart and lungs. As the heart rate and respiration signals are regular (periodical) and less noisy, the EMA and IPA can obtain clear results. It is noted that since the EMA is sensitive to noises, it is not suitable for analyzing noisy and nonstationary EEG recordings. In refs. [8,9], state-space and phase-dynamics approaches have been proposed to detect the weak coupling direction. The same drawbacks as those of EPA and IPA exist.A directionality index based on conditional mutual information is proposed and applied to the instantaneous phases of weakly coupled oscillators [10] , with which the coupling direction between oscillators can be identified. This method is abbreviated to IM in this paper. The adva...