Abstracts

Rationale: The identification of seizure generating tissue in patients with refractory epilepsy is crucial for successful neurosurgical intervention. Current methods are time-consuming and expert-dependent, with seizure freedom rates ranging from 40-80% across studies. High frequency oscillations (HFOs) are a promising biomarker of seizure generating tissue, but their specificity has fallen short in recent prospective studies. This could be due to reliance on an empirical definition of HFOs, which emphasizes the visual appearance in the filtered EEG; recent evidence has suggested that high frequency bands contain other information helpful for localizing epileptogenic channels. Therefore, we aimed to develop an automatic data-driven method to identify novel candidate biomarkers in the time-frequency (TF) domain to accurately delineate the seizure generating tissue. We apply this method to human iEEG and validate it using two independent datasets.


Methods: We detected high frequency (70-500 Hz) events of interest (EOIs) using our novel TF analysis and characterized them based on eight features - frequency of peak power, maximum power, area, duration, height, density, density of simultaneous events, and density of surrounding events. Each EOI’s features were categorized as “high” or “low” based on medians, using a leave-one-out method. EOI categories specific to resected channels were extracted from iEEG data from 9 refractory epilepsy patients from UCLA who were seizure free after surgery. Then, clustering was used to group similar categories together. We assessed the predictive power by classifying resected vs. non-resected channels using ROC curves. We validated the method on an open dataset from the Hospital of the University of Pennsylvania using the UCLA-derived categories and compared the performance to 5 HFO detectors.


Results: Two categories of events enable classification of resected and non-resected channels. Events in the first category, large broadband clusters (LBC), often resemble known electrographic events like HFOs, spikes, and HFOs on spikes (Fig.1B). LBCs had high maximum power ( >90 µV), area ( >1.04 Hz·s), duration ( >15.75 ms), height ( >50 Hz), and density of simultaneous events ( >0.074 AU). Events in the second category, small narrowband islands (SNI), had low frequency of peak power (70-210 Hz), area, height, and density of simultaneous events, and high maximum power, and duration. Visually, SNIs are distinct in the TF image but not in the 70-500 Hz filtered signal (Fig.1A). SNI and LBC events successfully identified resected channels (mean AUC = [0.72, 0.6], respectively) and largely outperformed 5 HFO detectors (mean AUC = [0.55, 0.52, 0.54, 0.51, 0.63], Fig.2). Validation in an independent dataset confirmed the association of these events with resected tissue in 10 seizure free subjects.


Conclusions: Our novel, data-driven approach identified two promising candidate biomarkers for seizure generating tissue, enabling classification accuracy that rivals or exceeds traditional HFOs.

Funding: National Institute of Neurological Disorders and Stroke of the NIH under Award Number R01NS116273 and UC Irvine California-Catalonia Engineering Program.