Abstracts

Rationale: The most effective treatment for medically resistant epilepsy (MRE) is surgical removal of the epileptogenic zone (EZ). A precise identification of the EZ is essential for surgical success, but success rates range from 20-80%. In patients with difficult-to-localize seizures, clinicians visually inspect intracranial EEG (iEEG) recordings during seizure events to localize the EZ. Although interictal data is also used, interictal discharges have not been proven to be reliable EZ markers. The need for seizure occurrence makes the process both costly and time-consuming and, in the end, less than 1% of the data captured is used to assist in EZ localization. In this study, we aim to leverage the interictal iEEG (ii-iEEG) data to localize the EZ. We hypothesize that in the interictal phase, the EZ is inhibited by adjacent cortical areas. We thus identify two groups of nodes from a patient's ii-iEEG network: those that inhibit a set of their neighboring nodes (“sources") and the nodes being inhibited (“sinks"). We then use source-sink (SS) connectivity properties of the network to identify hypothesized EZ nodes and predict surgical outcomes.

Methods: We used iEEG data from 31 MRE patients (14 successes, 17 failures) treated at the Cleveland Clinic. For each patient, we estimated a dynamical network model from the ii-iEEG data, which characterizes how the iEEG channels influence each other. We then placed the channels in a 2D SS-space and identified: a) sources, defined as channels that highly influence other nodes but are not highly influenced by others, and b) sinks, which are highly influenced by other nodes, but do not influence others. Next, we computed 3 indices for each channel to quantify its sink behavior and the amount of influence it receives from top sources and sinks. We expect all indices to be high in channels belonging to the clinically annotated EZ in success patients. Finally, we evaluated the SS-indices as EZ markers by building a logistic regression model to predict surgical outcomes.

Results: Fig. 1 compares SS-spaces of a success (A) and a failure (B) patient. In success patients, top sources point to top sinks and the sinks are highly connected. In this patient, all SS-indices are high. In contrast, top sources and sinks often point to other nodes in patients with failed outcomes, resulting in lower SS-indices. The logistic regression model was trained on 19 patients and tested on the remaining 12 using 10-fold cross-validation and yielded a 85.1+/-7.5% test set accuracy. Fig. 2 shows the ROC curve of the mean model (A) and the corresponding probability of success for each patient (B).

Conclusions: Our results suggest that channels that colocalize with the EZ have higher SS-indices compared to non-EZ channels. As such, the SS-indices may be promising markers of the EZ. Further, when used to predict surgical outcomes, the indices yield 85% test set accuracy (compared to a 45% success rate of the dataset). The SS algorithm could significantly improve surgical outcomes by increasing the precision of EZ localization.

Funding: Please list any funding that was received in support of this abstract.: AES Predoctoral Research Fellowship.