Abstracts

The tools used for quantitative analysis of the EEG have expanded to include methods used it physical sciences, including nonlinear dynamical systems measures. These tools have been used to 1) distinguish seizures from non-seizures, 2) perform seizure prediction and 3) to determine whether EEG segments include chaotic behavior. Despite the promise shown by these methods, a number of fundamental properties of the measures have not been described. The purpose of this paper is to describe the probability density functions for a number of nonlinear systems measures of dimensionality, entropy, global non-linearity and eigenvalues. We are especially interested in these functions as seen in the pediatric aged patient. Participants included both neurologically normal children (10) and those with epilepsy (17). Ten segments, five from sleep and five from awake, were procured from each subject. Each segment consisting of thirty seconds of EEG sampled at 400hz. The values were calculated for eigenvalue, Kolmogorov entropy, maximum likelihood correlation dimension, least squares correlation dimension and a Z score. The thirty second window was then shifted by one second and the values re-calculated (moving thirty times to produce a thirty second moving window). Data from each of ten electrodes were calculated (F3, F4, C3, C4, O1, O2, T7, T8, P7 and P8). Distributions were then constructed of 1) the absolute values of each of the non-linear metrics, 2) the number of consecutive samples changing in the same direction, 3)the magnitued of sample to smple change and 3) the percentage change from sample to sample. Measures of control tendency and variability were then calculated for each distribution and the distribution were fit with normal or gamma distributions. The trial to trial differences in consecutive values for normals were normally distributed for eigenvalue, dimensionality (both measures), entropy and the Z score with means all near zero (0.02, -0.125, 0, 0 and 0) with eigenvalue showing the narrowest distribution. The trial to trial change did not show a strong dependence on electrode location or state, but the distribution of the absolute values had a strong dependence on both variables. The distributions of absolute values for awake (0.56 and 0.75) and asleep (0.88 and 0.8) eigenvalues were bimodal and the distributions of entropy and Z during sleep were best fit by poisson distribution. The distributions of the number of samples showing n consecutive increases or decreases matched that expected if each consecutive sample were independent. The distributions of percent change from sample to sample were all normally distributed. There are significant differences in the shapes of the underlying distribution of the variables studied. This information must be taken into account when performing statistical comparison. (Supported by Falk Medical Trust Foundation)