Search

An unresolved question in epileptology is whether epileptic seizures can be predicted by characterizing measures of the EEG. In a recent comparison of univariate and bivariate measures we have shown the latter to be superior [Mormann et al., Clin Neurophysiol (submitted)]. However, so far only symmetric bivariate measures for synchronization were evaluated. In this study a comparison of 16 different measures for both synchronization and directionality was carried out with respect to their capability to discriminate pre-ictal from inter-ictal intervals with special emphasis placed on statistical validation [Kreuz et al., Phys Rev E, 2004]. Furthermore, we investigated to which extent different measures carry independent and non-redundant information. We analyzed continuous multi-day EEG recorded intracranially from 9 patients (total duration: 860 hours, 66 seizures). Measures included symmetric ones like cross correlation, mutual information and phase synchronization (based on Hilbert and wavelet transform, respectively) as well as asymmetric ones like transfer entropy, non-linear interdependencies and event synchronization. Measure profiles were evaluated using a statistical approach based on Receiver-Operating-Characteristics (ROC). The analysis was performed as in our previous work [Mormann et al., Clin Neurophysiol (submitted)] allowing different lengths of the pre-ictal interval and testing for both a pre-ictal decrease and increase. Two different evaluation schemes were designed focussing either on global or local effects. The significance was estimated using seizure times surrogates [Andrzejak et al., Phys Rev E, 2003]. Finally, redundancies between these measures were quantified using both correlation and cluster analysis. In the first evaluation scheme all measures yielded very poor ROC values and no measure proved to be significant. In the second evaluation scheme performances were considerably higher and, more importantly, for the most part significant. Maximum values were obtained for the indices of phase synchronization, minimum values for the symmetrized versions of the non-linear interdependencies. While the measures of directionality covered the same range as the measures of synchronization, correlation analysis showed that both groups of measures carry independent information. While no significant global effect could be observed, some measures of synchronization as well as some measures of directionality proved capable to detect significant local effects. Due to their non-redundancy a combination of these different approaches appears very promising. However, much more research has to be carried out to evaluate whether these statistical results allow yielding an acceptable performance in a practical implementation.