Meanwhile, numerous publications address the feature extraction problem in epilepsy. Up to now a precursor detection based on changes of EEG-signal features could not be performed with a sufficient sensitivity and specifity for an automated seizure warning system. Different approaches including procedures using stochastic models, as well as algorithms based on Cellular Nonlinear Networks (CNN) and Volterra-Systems have been discussed throughout previous publications. Therm interesting findings have been discussed involving e.g. signal prediction algorithms and the calculation of synchronisation measures. In this contribution new results obtained in a spatio temporal linear prediction of segmented electrode signals using long-term SEEG and ECoG recordings of patients in epilepsy will be discussed in detail.