Towards an automated seizure anticipation device based on cellular neural networks (cnn)

The epileptic disorder, already mentioned in a Babylonian text dated from the middle of the first millenium BC, nowadays is known to be the most common chronical disorder of the nervous system. Epileptic seizures are phenonema of abnormal synchronization of neural activity with symptoms like convulsions and generally strike without warning. Epileptic drugs, taken in the most important therapy, have the disadvantage of adverse effects and possible habituation. A reliably, automated seizure warning system would not only provide valuable information to the patient, but also enable an efficient, event specific therapy. The problem of detecting a possible pre-seizure state in epilepsy from electroencephalogram (EEG) signals, has been addressed by many authors over the past decades but still remains unsolved. Provided that the transition between interictal state and the ictal event is not an abrupt phenomenon but a gradual change in dynamics [1], [2], precursors could be detected by analyzing brain electrical activity. Several publications report evidence that a preictal state can be detected in focal epilepsy by considering multi-variate measures [3], [4], [5], [6], [7], [8], [9] in particular, although seizures cannot be anticipated with necessary sensitivity and specificity up to now. In this contribution models based on CNN are considered in order to analyze signals from intracranial EEG, taking into account mutual dependencies between signals of neighboring electrodes. Due to their inherently parallel paradigm of computation and their high processing speed under real-time conditions combined with low power consumption, CNN are well suited to a great extent for the processing of multi-dimensional bio-electrical activity and a promising candidate for a future implantable seizure warning and preventing device. In the first proposed algorithm, solutions of Reaction-Diffusion CNN (RD-CNN) models are used in order to approximate short segments of EEG-signals. In a second algorithm, the behavior of linear spatio-temporal systems represented by discrete-time CNN (DT-CNN), are used for signal prediction of intracranial EEG. Results for the analysis of long-time recordings gained during presurgical diagnostics in temporal lobe epilepsy are given regardimg both algorithms and their predictive performance with respect to impending epileptic seizures is evaluated statistically. Additionally, the second above mentioned algorithm has been implemented on the Eyes-RIS 1.1 system [10], [11]. First results for the analysis of intracranial long-time recordings carried out on this system are given.