In different investigations it has been shown that nonlinear signal processing can contribute to the task of finding precursors of impending epileptic seizures in the case of a focal epilepsy [6]. Various approaches to this feature extraction problem have been made including Volterra-Systems [2],Wavelet-Analysis and Cellular Neural Networks (CNN) [3]. This paper gives a detailed analysis of a recently proposed prediction [1] algorithm based on multi-layer delay-time DTCNN. The aim of this contribution is to reduce the high computation complexity caused by the permanent application of a supervised optimization procedure for successive data segments of an EEG recording. Thereby the prediction algorithm is studied by using different optimization procedures, different network topologies and different template symmetries.