Abstracts

Anticipation of epileptic seizures is, among others, the most challenging aspect in epileptology. Recent studies have shown that particularly measures which characterize the degree of synchronization between two EEG signals allow an improved differentiation between the seizure-free interval and the pre-seizure period. Despite the conceptual simplicity of a number of these bivariate measures, real-time applications are currently limited by calculations for large number of combinations of electrodes. In this study we examine the ability of Cellular Neural Networks (CNN) to accurately approximate the degree of synchronization (as defined by the mean phase coherence R) between two EEG signals. CNN have a massive computing power, are capable of universal computation, and are already available as analog integrated circuits. In order to find optimum network settings we performed an in-sample supervised training using 24 randomly selected pairs of EEG epochs (epoch duration: 23.6 sec.) recorded intracranially during the seizure-free interval and the pre-seizure period from three epilepsy patients along with the corresponding values of the mean phase coherence R. For an out-of-sample validation and in order to study the long-term behavior of our CNN-based approximation for phase synchronization, network settings were then tested on multi-day, multi-channel EEG recordings. CNN with polynomial weights allowed to approximate the temporal evolution of mean phase coherence R with a stability and an accuracy (more than 90%) that can be regarded sufficient to distinguish between the seizure-free interval and the pre-seizure period. Interestingly, the obtained CNN allowed to approximate R values even from the ictal and post-ictal state with a comparable accuracy although we did not use data from these states for the training. CNN allow an accurate and stable approximation of the degree of phase synchronization between two EEG signals. This ability along with the high computational power and the small energy and space requirements render CNN attractive for future implementations in a hardware environment to be used as a miniaturized supercomputing device for real-time applications. (Supported by The Deutsche Forschungsgemeinschaft)