Abstracts

Many algorithms have been used for the automated detection of seizures. Most recently, methods based upon nonlinear dynamic systems analysis have been used, generally with small data sets in adult patients. The purpose of this study is to describe the application of adaptations of nonlinear dynamic systems measures of eigenvalue, Kolmogorov Entropy (KE), two forms of correlation dimension, and a global measure of nonlinearity (z) to the detection of seizures in pediatric patients. Reliable estimates of these measures require large data sets. At conventional sampling frequencies this requires very long epochs, thus compromising temporal resolution. We introduce the use of a [quot]moving window[quot] version of the above dependent variables and the application of a statistical test for the recognition of a significant change in the EEG signal.
Data sets were obtained from pediatric patients according to an IRB approved protocol. Three thirty second artifact free epochs, from 100 studies, were selected for both sleep and wake states and compared to the segments containing the seizures. Each of the variables were measured on the data sets and statistical criteria (variation of the sign test) used to determine [quot]significant[quot] changes in the EEG. Data epoch length was fixed at 30 seconds but was generated using a moving window in one second steps. Each single thirty second moving window result was then compared to a longer duration baseline measure to determine whether or when there was a consistent change.
The moving window implementation detected more than 80% of the seizures. The most stable measure was the eigenvalue while the least stable was the least squares implementation of the correlation dimension. For most of the dynamical systems measures the inclusion of only one or two seconds of seizure, out of thirty seconds , produced a substantial change in the value for the entire segment. In a number of cases there were changes in the the dynamical measures prior to the visually determined onset of the seizures (seizure anticipation). The eigenvalue was the least sensitive in terms of anticipation while the maximum likelihood correlation dimension and the KE were the most sensitive. Finally, the application of a statistical criteria dependent on the deviation of a series of points from the baseline proved less vulnerable to artifactual false positives than single point standard deviation rules.
Dynamical systems measures can detect seizures in the pediatric patient . A method for statistically determining when an EEG segment is different from its baseline is given and it compared favorably to alternative measures. There are significant differences amongst the [quot]chaos[quot] measures, in terms of stability, sensitivity and computational load. We also noted that it was possible to [quot]anticipate[quot] nearly 50% of the seizures within thirty seconds of onset of the seizure.
[Supported by: Falk Medical Trust foundation]