Abstracts

Rationale: Epilepsy remains a prominent and costly disorder making research towards the development of antiepileptogenic treatments a high priority. Because of its latent, multifactorial nature, little is known about the processes that underly epileptogenesis and early diagnosis of patients undergoing epileptogenesis is a problem yet to be solved. Discovery of a biomarker of epileptogenesis and subsequent curation of accurate experimental and test groups is a critical step towards justifying clinical tests of novel antiepileptogenic treatments.One potential biomarker is high-frequency oscillations (HFOs) in EEG (Annals of neurology. 2012 Feb;71(2):169-78). Although HFOs may hold promise, different definitions exist in literature and automatic detectors often detect spike activity or noise in addition to true HFOs. The goal of this study is to analyze large EEG datasets and find a meaningful clustering of HFO detections from automatic software to provide insight into detecting epileptogenesis. Methods: Sixty hours of scalp EEG data sampled at 2048 Hz were recorded from five patients who had recently suffered moderate to severe traumatic brain injury (TBI) from Massachusetts General Hospital, Phoenix Children’s Hospital, and University of California, Davis. HFO detection algorithms have typically focused on intracranial EEG as opposed to scalp EEG due to the amount of noise found in the latter. In addition to being recorded farther from the brain, scalp EEG data are noisy because electrical potentials scatter laterally upon contact with the skull, delocalizing electrode information. The surface Laplacian, a nonlinear spatial filter, was applied to minimize these volume conduction effects and increase topographical localization as seen in Figure 1 (processed data on the left and raw data on the right). The application of a notch filter and a Surface Laplacian, which is less prone to error compared with discrete low/high pass filters was applied. The data were band-pass filtered from 80-500 Hz, and after filtering, segments with a root-mean-square energy 5 standard deviations above the baseline energy lasting at least 4 oscillations were detected as possible HFOs. Finally, Unsupervised Diffusion Component Analysis (UDCA) was applied to reduce the dimensionality of the data and reveal the underlying brain activity by meaningful geometric patterns (Math Biosci Eng. 2016 Dec 1;13(6):1119-30). Based on the diffusion mapping framework, UDCA is a manifold learning technique adapted to fit the noisy, stochastic nature of EEG. Results: Figure 2 shows an embedding using the two eigenvectors that displayed the greatest variability within the data. 79.94% of the 6384 detected events (average 55.45ms) were embedded near the origin and subsequently compared to clusters of events considered outliers. Each point in the embedding represents one detected event, and each color represents a different subject. The embedded points followed a similar pattern for each subject, and those points farthest away from the origin represented detections that appeared to be closer to true HFOs rather than spikes or other epileptiform activity detected by the software. Conclusions: Detection and classification of HFOs has become increasingly automated. The preliminary result of this study is a fully automatic tool that can efficiently analyze large amounts of data, detect potentially pathogenic waveforms, and arrange the events in a lower-dimensional space based on their geometric stucture. Future work will include visual classification of waveforms followed by statistical testing to determine the accuracy of the algorithm at detecting HFOs. Development of this algorithm into a fast and reliable predictor of epileptogenesis would be critical for testing potential treatments that prevent post-traumatic epilepsy and to develop a model to understand epileptogenesis. Funding: NIH U54NS100064 (EpiBioS4Rx)