Abstracts

Interobserver Agreement of Seizure Detection Using Supervised Computation EEG Analysis

Abstract number : 1.149
Submission category : 3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year : 2016
Submission ID : 193873
Source : www.aesnet.org
Presentation date : 12/3/2016 12:00:00 AM
Published date : Nov 21, 2016, 18:00 PM

Authors :
Brittany Van Nelson, Harvard Medical School, MGH, Allston, Massachusetts; Kimberly Lu, Harvard Medical School, MGH; Humzah Mahmood, Harvard Medical School, MGH; Theju Jacob, Harvard Medical School, MGH, Charlestown; F. Edward Dudek, University of Utah Sch

Rationale: Detecting the difference between artifacts and ictal events can be challenging for even the most seasoned EEG expert [1]. The objective of this study is to assess the reliability of seizure detection using supervised computer algorithms in multi-channel EEG data covering thousands of hours of EEG recordings from chronic in-vivo experimentation. Methods: Ninety-four, eight hour EEG recordings from one hundred different rodent biomarkers were analyzed in this study. The first round, Observer 1 was unblinded to the files names which indicated the rat of origin, the date of epileptogenic brain injury, and the time elapsed since injury. Seizures were detected using the algorithm described below. The recordings were then randomized, and Observer 2 was blinded to the rodent biomarker file names, and seizure incidence was noted in single-channel files. Ictal event candidates were identified with DClamp (freeware: https://sites.google.com/site/dclampsoftware/home). This program is a data acquisition and analysis software for epilepsy research that provides automatic detection of ictal-like events and interictal spikes in EEG recordings. Upon completion, both data sets were unblinded and seizure incidence results were compared. Interobserver agreement was calculated from the expected and observed agreements to compute the kappa coefficient (K ). Results: The unblinded and blinded analyses agreed on eighty-seven out of ninety-four biomarker files. Since the K value was above 0.81 (K 0.86, 95% CI 0.77-0.95), there was substantial interobserver agreement of seizure detection between unblinded and blinded files analysis [2]. Conclusions: Blinding observers as to the origin of the EEG data is an important strategy for increasing the robustness of preclinical studies, but may result in reduced accuracy of seizure detection. We found that blinded and unblinded interobserver agreement was very good when detecting seizures via supervised computer EEG analysis. We conclude that supervised computational analysis of blinded EEG data is a feasible strategy for seizure detection in large, experimental EEG data sets. Funding: 5R01NS086364-03 (Staley) 09/01/13-05/31/18 NIH/NINDS
Neurophysiology