Abstracts

Rationale: The best possible way to achieve complete seizure control for patients with drug-resistant epilepsy is through surgically resecting the seizure onset zone (SOZ), the brain area where the abnormal, excessive neuronal activity starts. We believe that the key to addressing the problem of SOZ localization is to understand the directional connections among brain regions. First, the human brain is a directional network system, because each brain region constantly exerts influence over or passes information to other regions through neuron firing. This directional influence exerted by neuronal activities from one region to others is commonly referred to as effective connectivity. Second, most importantly, the seizure propagation, which originates from the SOZ(s), is directional and dynamic. Thus, understanding changes in the brain activity and directionality of neuronal information passed from one region to others is critically important to epilepsy diagnosis. Methods: We build a novel high-dimensional dynamic model for the brain’s directional network using intracranial EEG (ECoG) data, where each network node corresponds to a brain region recorded by ECoG, and each directional network edge indicates a directional effect exerted by one region over another. In contrast to existing complex dynamic models that rely heavily on the strong, specific knowledge of the brain regions under study, the proposed high-dimensional model is widely applicable to characterize various brain regions' oscillatory activity and to explore their connectivity patterns. We then develop a unified Bayesian framework to estimate the proposed model, identify strongly connected brain regions, and map the brain’s directional network. Results: Through simulation studies, we found that popular correlation and partial correlation measures are not as effective as the proposed dynamic model for identifying strongly connected brain regions. We applied the developed Bayesian method to 1-second ECoG data segments collected from an epileptic patient whose SOZ is in the temporal lobe. We found that the brain regions in the temporal lobe are most strongly connected. However, the SOZ is disconnected from the other regions in the segments close to the seizure onset.  Conclusions: We argue that the proposed model is more useful for evaluating connections among brain regions than correlation measures in practice for two reasons. First, the correlation and partial correlation essentially measure the linear relationship among the brain activity at different regions, and they do not directly accommodate the brain’s nonlinear periodic activity. Second, the two correlations measure synchronized activity of brain regions, while the proposed model directly characterizes how each region’s activity temporally affects other regions’ activities. Overall, the new model characterizes more complex properties of the brain activity than the simple correlation measures. The observation of isolated SOZ from the rest of regions before the seizure onset is in ine with existing network results. We also found that this result holds only when the seizures start from only one SOZ. This is possibly because when there are multiple SOZs, the connections between SOZs and between SOZs with other normal regions are too complex to identify a simple pattern.  We believe that the unique connectivity property of the SOZ identified by the proposed method can be potentially used to help clinicians to locate the SOZ in practice.  Funding: The proposed research was partially supported by University of Virginia Quantitative Collaborative Seed Grant. Award.