Authors :
Presenting Author: David Adams, MD – Children's Hospital of Orange County
Derek Hu, PhD – Biomedical Engineering – University of California, Irvine; 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 – UCLA Mattel Children's Hospital; Beth Lopour, PhD – University of California, Irvine
Rationale:
The identification of various EEG waveforms, such as slow spike-and-wave and generalized paroxysmal fast activity, is critical for diagnosing Lennox-Gastaut Syndrome (LGS) but is complicated by the progressive nature of the disease. Here, we assess the interrater reliability (IRR) among pediatric epileptologists for classifying EEG waveforms associated with LGS.
Methods:
A novel automated algorithm objectively selected 1,350 EEG events from 20 LGS subjects and 20 approximately age-matched healthy controls during NREM sleep. Three raters independently reviewed the events within isolated 15-second EEG clips in a randomized, blinded fashion. Each event was labeled based on the subject type (i.e., healthy control or LGS subject), and the event type (i.e., spike and slow-wave, generalized paroxysmal fast activity, seizure, sleep spindle, vertex sharp, muscle, artifact, other event, or nothing). IRR between clinicians was quantified using Cohen's kappa and the rater agreement was calculated for each event type.
Results:
Labeling of subject type had 85% accuracy across all events and an IRR of κ = 0.790; labeling of individual EEG events had κ = 0.558. The event agreement rate for subject type was 80.6% for controls and 81.3% for LGS subjects. For event type, the event agreement rate was highly variable for different events: 71.0% for spike and slow-waves, 28.5% for generalized paroxysmal fast activity, 49.0% for sleep spindles, and 12.2% for vertex sharps.
Conclusions:
Brief epochs of EEG containing high power events can be reliably classified as pathological or normal. The event classification was less consistent between clinicians; mismatches most frequently occurred when one reviewer selected “no event,” suggesting that the threshold for labeling an event varied between reviewers. Computational methods can help objectively define these events and may improve IRR and aid clinical decision making. In the long term, facilitating these automated algorithms can quantitatively define waveforms as objective biomarkers, improving the diagnosis and treatment of patients with LGS.
Funding:
This work was supported in part by the Lennox-Gastaut Syndrome Foundation and the John C. Hench Foundation.