Abstracts

AN AUTOMATED SEIZURE WARNING ALGORITHM FOR SCALP EEG

Abstract number : 2.155
Submission category :
Year : 2003
Submission ID : 3685
Source : www.aesnet.org
Presentation date : 12/6/2003 12:00:00 AM
Published date : Dec 1, 2003, 06:00 AM

Authors :
James C. Sackellares, Leon D. Iasemidis, Deng-Shan Shiau, Wachai Suharitdamrong, Linda K. Dance, Wanpracha Chaovalitwongse, Panos M. Pardalos, Paul R. Carney Neurology, University of Florida, Gainesville, FL; Psychiatry, University of Florida, Gainesville

We recently developed the first automated prospective seizure warning algorithm (ASWA) for intracranial EEG, which predicts impending seizures with an average sensitivity of 84%, specificity of 0.12 false predictions per hour, and warning time of 72 minutes prior to a seizure onset ([italic]Iasemidis et al., IEEE Trans. Biomed. Engin.,50, 616-627, 2003).[/italic] We herein report prediction results from the application of our algorithm to long-term scalp EEG recordings.
Continuous 22-channel, long-term (6.7 to 10 days) scalp EEG recordings previously obtained from 3 patients with medically intractable partial seizures was analyzed by ASWA. Patient 1 had 11 seizures during a 6.7 days recording (mean inter-seizure interval MII=7.2 h); patient 2 had 11 seizures during 8.9 days recording (MII = 10.6 h), and patient 3 had 14 seizures during 10 days recording (MII = 14.5 h). The seizure warning algorithm involved the following steps: (1) calculating the short-term Lyapunov exponent ([italic]STLmax[/italic]), a measure of stability of dynamical systems, for sequential 10.24 second non-overlapping epochs from each electrode site, (2) calculating paired-T statistics (statistical distance of mean [italic]STLmax[/italic] values) in 10-minute windows before and after the previous seizure for all possible electrode pairs, (3) selecting the critical group of electrode sites using integer quadratic optimization, (4) calculating the average T-statistic among the selected critical electrode sites (T-index) forward in time and (5) warn of an impending seizure when observing a dynamical entrainment transition (convergence of critical sites[rsquo] [italic]STLmax[/italic], mean difference = 0 at a=0.01 significance level). The warning was considered to be correct if a seizure occurred within 3 hours thereafter and false if it did not.
The prediction sensitivity of this algorithm was 80.0%, 80.0%, and 92.3% for patients 1, 2 and 3 respectively (84.9% overall). The false prediction rates were 0.236, 0.312, and 0.280 per hour (0.281 / h overall). The average warning times before seizures were 79.9, 93.5, and 85.4 min for patients 1, 2 and 3 respectively (85.70 min overall).
Our results indicate that this prospective seizure warning algorithm appears to be capable of warning of an impending seizure by analyzing standard scalp EEG recordings. In this test sample, the algorithm performs with sensitivity and false positive characteristics that indicates that it may have practical clinical utility. For example, the algorithm could be incorporated in recording systems used in an epilepsy monitoring and intensive care units.
[Supported by: NIH/NIBIB 8R01EB002089-03
U.S. Veterans Affairs
The Whitaker Foundation
NSF]