Abstracts

Rationale: Precise identification of the epileptogenic zone (EZ) is essential for presurgical planning and can be problematic when based on inadequately localizing intracranial electroencephalography (iEEG) data.  Our prior studies suggest that changes in interictal brain network connectivity occur consistently within 1 minute prior to seizure onset, beginning in EZ, and might be used as an independent identifier of regions critical for seizure generation to help ensure optimal surgical outcomes. In this study, we develop a decision support system for automated identification of EZ using effective connectivity measures estimated from multiple input, single output (MISO) state space models based on the preictal iEEG.

Methods: We analyzed iEEG from 10 patients selected from the iEEG.org database who were seizure-free post-surgery (ILAE Class I) and had clearly defined and spatially restricted seizures. Recordings for each patient were common mode average referenced and bandpass filtered (0.5-250 Hz).  For each seizure, the 2 minutes prior to seizure onset were divided into 5 second epochs. In each window, a multiple input, single output (MISO) state space model was estimated for each channel output with all other channels as inputs. The magnitude of the parameters describing the influence of each other channel on the given channel were used to infer a directed network graph of the relationship between all channels for the duration of the window. Degree centrality was calculated for each channel and was used as the feature for training a radial basis function support vector machine (SVM) classifier using 5-fold leave-one-out cross-validation. Classifier success was assessed by True Positive Rate (TPR) and Accuracy (ACC) at the seizure, patient, and population levels, using the clinically-identified EZ electrodes as gold standard.

Results: TPR and ACC values are shown (Figures 1, 2) for all 10 patients averaged across seizures, as well as the average across patients. The MISO degree SVM classifier achieved a minimum TPR and ACC in all patients of .85 and .89, respectively, with an average TPR of .90±.03 and ACC of .95±.04. The classifier correctly identified the entire EZ in 73% of seizures, missed 1 channel in 21% of seizures, and missed 2 channels in the remaining 6% of seizures. No channels were identified by the SVM classifier that were not clinically identified as EZ channels.

Conclusions: Changes in effective network connectivity during the pre-ictal period are sufficiently distinct to allow for highly accurate identification of the EZ using simple machine learning tools. Since these network changes are observed before the electrographic seizure, they constitute an independent confirmation of specific EZ regions in epileptogenesis. This study provides novel supporting evidence for the critical contribution of brain networks in focal seizure generation and the instigating role of EZ at seizure onset. The high reliability of this decision support system offers a new computer-assisted tool for resective surgical planning, and may be of particular value when localization based on standard iEEG is less clear.

Funding: UConn SOM, CT Institute for Brain and Cognitive Sciences