Abstracts

An important issue in epileptology is whether epileptic seizures can be anticipated prior to their occurrence. Of particular interest is the question whether information extracted from the EEG of epilepsy patients can be used for the prediction of seizures. Several studies have claimed evidence for the existence of a pre-seizure state that can be detected using linear and nonlinear EEG analysis methods. Most of these studies, however, were performed on short, selected recordings and to date there is little experience with continuous long-term recordings over several days. In this study we evaluate the predictive performance of a variety of measures derived from the theory of dynamical systems.
We compare different linear and nonlinear measures comprising both univariate (derived from a single EEG signal) and bivariate approaches (characterizing the synchronization between two EEG signals recorded simultaneously from different locations of the brain) in terms of their ability to distinguish between the seizure-free interval and the pre-seizure period. We analyze intracranial continuous multi-day EEG recordings from up to now 5 patients with temporal mesial lobe epilepsy undergoing presurgical diagnostics covering more than 300 hours of multi-channel EEG. We use Receiver-Operating-Characteristics (ROC curves) to quantify the degree to which the amplitude distribution of the time profiles calculated from the seizure-free interval can be distinguished from the amplitude distribution from the pre-seizure period for a particular measure. In order to assess the statistical significance of the obtained performance values, we use a recently proposed method termed seizure time surrogates.
Our analysis shows a similar performance of linear and nonlinear measures. We do, however, find a distinct difference between bivariate and univariate approaches with a higher performance of the bivariate measures. Validation analysis using seizure time surrogates shows only performance values for the bivariate measures to be statistically significant.
The predictive performance of univariate measures reported in earlier publications could not be confirmed. Results for the bivariate measures, on the other hand, provide statistically significant evidence for the existence of a pre-seizure state that is reflected by changes in synchronization as characterized by these measures.
[Supported by: The Deutsche Forschungsgemeinschaft.]