Abstracts

Rationale: High frequency oscillations (HFOs) are a biomarker of epileptic tissue, being developed to help guide resective surgery. One major challenge limiting clinical translation is that HFOs can occur due to both normal and epileptic processes. One potential method to distinguish between normal and epileptic HFOs is to compare "ictal-like" versus "interictal-like" HFOs. We hypothesize that healthy tissue can produce ictal-like HFOs, but that this should only occur when seizures have spread to that tissue. In contrast, we hypothesize that epileptic tissue may have "ictal-like" HFOs at any time. Methods: In order to classify HFOs, we used a validated automated algorithm in intracranial EEG recorded from 17 patients with refractory epilepsy. For each of the resulting >1.6 million HFO detections, 33 features were computed, including duration, power, line length, etc. The topology of the feature space was assessed using a non-parametric estimate of intrinsic dimension, and changes in the topology relative to time and channel were quantified using angular distance and a generalized Grassman-chordal distance. Linear discriminant analysis was used per channel to reduce the dimension, followed by a support vector machine (SVM) to classify the HFOs. This process was used to classify whether each HFO had ictal features. We then limited analysis to interictal periods and considered the correlation in number of "ictal-like" HFOs versus seizure onset, quantified using the count asymmetry. Results: The specific feature manifold determined by the linear discriminant analysis was found to vary between channels and patients. However, the asymmetry in HFO rates with respect to the seizure onset zone improved when using only "ictal-like" interictal events, rather than all interictal HFOs. Conclusions: These methods demonstrate a novel, effective method to classify abnormal HFOs, which can have significant impact in their translation as epilepsy biomarkers. Funding: We gratefully acknowledge funding from the Doris Duke Foundation (WCS), US National Institute of Health (NIH) grant K08-NS069783 & R01-NS094339 (WCS), NIH Big Data to Knowledge Mentored Training Grant K01-ES026839 (SVG), US National Science Foundation (NSF) grant CCF-1217880 (KRM,AOH), and an NSF Graduate Research Fellowship, grant F031543 (KRM).