Abstracts

Despite a number of approaches that have been introduced to the field of seizure prediction the detection of features predictive of an impending seizure remains unsolved. Recent studies have shown that nonlinear feature extraction algorithms may be useful for the detection of precursors that exhibit distinct changes in the pre-seizure period. In previous investigations we have proposed algorithms for signal prediction based on multi-layer discrete time Cellular Neural Networks (DTCNN) and Volterra systems. Furthermore, we introduced a so called Pattern Detection Algorithm (PDA) for standard CNN which is based on the analysis of the level-crossing behavior of the EEG. In this contribution recent results for the CNN prediction algorithm and for the PDA will be discussed in detail. Firstly, PDA is based on a binary input-output CNN with linear weight functions. A CNN with 72x72 cells is used to map an EEG segment of 30 seconds duration (5184 data points) to the CNN output activity. Depending on the neighborhood range of the considered network a so called window pattern is defined, e.g. as a 3x3 binary pattern, which is obtained from the binarized EEG. As will be shown later, one or more patterns, that show distinct changes in their rate of occurrence before seizure onset, can be found for each patient. Secondly, based on multi-layer DTCNN with polynomial weight functions, the nonlinear prediction of the EEG is discussed. The coupling weights and the prediction error of the obtained predictors have been analyzed in order to find appropriate features which may be used for a pre-seizure state identification. Using the PDA our results show that for each patient one or more patterns can be found, showing distinct changes in their rate of occurrence before seizure onset. Two types of patterns can be found for all patients that are defined by their type of occurrence. A first overview of the results obtained from a nonlinear prediction of the EEG using DTCNN shows that the prediction error and a measure derived from the coupling weights exhibit distinct changes before and at seizure onset. Applying these techniques to long-term EEG recordings we observe a decrease of the linear weights over a period of several hours which is accompanied by an increase of the nonlinear weights. Our preliminary results underline the importance of CNN-based feature detection algorithms for a realization of a miniaturized seizure prediction device. CNN-based analog cellular computing is a unified paradigm for universal spatial-temporal computation. CNN-based realizations with stored programmability show an enormous computation capacity with trillions of operations on a single chip. (Supported by the Deutsche Forschungsgemeinschaft)