Abstracts

Rationale: The challenge of epilepsy surgery is to identify and safely remove the causal part of an epileptogenic network. Intracranial EEG (iEEG) data, particularly ictal data, is currently sought in many cases to accomplish this purpose. When iEEG is used, current practice is to wait for clinical seizures, and visually assess ictal recordings. Our rationale for this study is that the networks supporting seizure generation are present whether a seizure is happening at a given moment or not, which raises the possibility that causal networks might be detected in interictal iEEG. Visualization of such networks may dramatically accelerate identification of epileptogenic regions. Methods: To visualize causal connectivity, multivariate Granger causality analysis was applied to interictal iEEG recorded from 25 epilepsy patients who underwent invasive monitoring for surgical resection planning at our institution. For each patient, all possible pairwise interactions were evaluated for directional causality as defined by Granger (Granger, C.W.J. 2013. Prize Lecture: Time Series Analysis, Cointegration, and Applications. Nobel Media AB). For each case, we analyzed 20 minutes of interictal iEEG data. To evaluate whether the causal map made from interictal data is correlated with the ultimate seizure onset localization made from the review of ictal data by the clinical team, we assessed the rank order sum of the specific electrodes determined by the algorithm in comparison with the seizure onset zone defined clinically. We also evaluated each causal map against the final resection area. To test statistical significance, null distributions were generated based on rank order sum and minimum distance to the resected area. Results: With the significance level of 0.05, p values were calculated using rank order sum test. Cases showing lower than 0.05 consistently reflect similarity between the electrodes predicted as causally influential set and those clinically identified as seizure onset zone. The statistical significance has been found in 17 out of 25 cases (eight with p< 0.005; as low as< 0.00001). For the entire data set, the overall p value is less than 10^-8. In addition to rank order sum test, we obtained expected distance from empirically created null distribution of distances between randomly selected electrode set and resected set, and used two-sample t-test to test statistical significance. At the 0.05 level, the population means of the distances to the resected area from the predicted set versus from the randomly selected set is significantly different from zero (p = 0.00008). Conclusions: Identification of the "high causality" regions of the brain from interictal iEEG data statistically predicts both the seizure onset zone and the resection area, which were independently based on ictal data. The result suggests an approach to finding the seizure onset or resection zone without the risk of long term intracranial monitoring, with its expenses and risks. This work is funded in part by NIH (1R01NS069696) and in part by Wyss Institute for network data storage and management.