Abstracts

RATIONALE:_Several methods for seizure prediction in EEG have been evaluated in the past. They were based on the calculation of parameters which characterize the non-linear dynamics of the EEG signal like correlation dimension or maximum Lyapunov coefficient. An important disadvantage of these methods is the enormous computational afford. Therefore we developed a method based on repetitive short predictions of the EEG with Elman-networks. Our decision to use this type of artificial neural network is based on the possibility of real-time computation of seizure prediction in the future. METHODS: The EEG signal was recorded from foramen ovale electrodes during presurgical video EEG monitoring of patients suffering from a pharmacoresistant epilepsy. For prediction of the EEG artificial neural networks with recurrent links (Elman-networks) were used. It is known that the maximum Lyapunov coefficient and the Kolmogorov entropy drop minutes before a seizure evolves in the EEG. The reciprocal of the Kolmogorov entropy can be understood as the average predictability of a time series. Therefore the prediction error of the EEG prognosis should drop before seizure onset. Up to now we evaluated three seizures of patients with an epilepsy of mesio temporal origin. In each seizure 60 minutes of preictal EEG were analyzed. In each minute one second (200 data points) of EEG were used for training and the prediction task. Stationarity was assesed by visual inspection of the EEG sections. Training was performed five times with randomly initialized parameters for each EEG segment. The network with the best training result was used to predict the EEG. The difference between the EEG signal and the prediction was calculated. RESULTS: In two seizures a significant reduction of the average prediction error could be found about 10 minutes before seizure onset. In one seizure no changes were seen. Maybe this result was due to the short EEG section analyzed in the patient. CONCLUSIONS: Our preliminary results suggest that the use of repetitive short predictions of EEG by Elman-networks may be an effective method for seizure prediction in epileptic patients.