Authors :
Presenting Author: Derek Hu, PhD – California State University, Long Beach
Marco Pinto-Orellana, PhD – Biomedical Engineering – University of California, Irvine; David Adams, MD – Division of Neurology – Children's Hospital of Orange County; Mandeep Rana, MD – Division of Neurology – Children's Hospital of Orange County; Linda Do, IMG/EFCMG – Children's Hospital of Orange County; Daniel Shrey, MD – Pediatric Neurology & Epilepsy – Children's Hospital of Orange County; Shaun Hussain, MD – Associate Professor, Pediatric Neurology, UCLA Mattel Children's Hospital; Beth Lopour, PhD – Associate Professor, Biomedical Engineering, University of California, Irvine
Rationale:
The development of EEG biomarkers for epilepsy often relies on observations from visual review, however, this empirical approach has significant limitations. For example, generalized paroxysmal fast activity (GPFA) is an emerging biomarker for Lennox-Gastaut Syndrome (LGS), but the lack of a precise physiological definition for GPFA limits its use and hinders the development of automated detection algorithms. Therefore, we developed a novel method for unsupervised time-frequency image analysis to automatically identify and characterize potential EEG biomarkers without relying on empirical clinical definitions.
Methods:
A 10-minute clip of scalp EEG during NREM sleep was retrospectively collected for 20 subjects with LGS and 20 approximately age-matched healthy controls. Continuous time-frequency points with power exceeding the baseline were marked as events. Each event was characterized based on frequency range, duration, spatial spread, time-frequency density, frequency of peak power, and mean power in the delta, theta, alpha, sigma, beta, and gamma frequency bands. Similar events were grouped together based on their features using two different clustering methods. To aid interpretation of the results, three blinded raters independently classified a subset of 1,350 events selected equally from each cluster.
Results:
A total of 11,708 events were identified across 40 subjects consisting of 4,964 LGS events and 6,744 control events. Six of the 12 clusters had a significantly greater number of LGS events than control events, making them candidate biomarkers of LGS (p< 0.05, Permutation Test, Bonferroni corrected). Visual review revealed that four of these clusters primarily contained known LGS-associated waveforms, such as GPFA and trains of interictal epileptiform discharges (IEDs). Moreover, there were only five sleep spindles found in LGS subjects, compared to 113 in healthy controls. An automated spindle detector supported this finding and reported a significantly lower count, duration, spatial spread, and frequency of peak power in LGS spindles compared to healthy controls (p< 0.05, Mann-Whitney U-Test). The last two candidate clusters consisted of short, high-frequency events that were labelled by raters as “nothing,” but 81.6% of these events came from LGS subjects. A second clustering method corroborated this finding, showing that 81.3% of events with high beta/gamma power were from LGS subjects.