Multivariate spectral estimation based on parametric modelling has been applied to epileptic surface EEG in order to detect EEG changes that occur prior to the clinical outbreak of the seizure. A better time/frequency resolution has been achieved using residual energy ratios (Dickinson's method). Prediction of oncoming seizures was based on detection of increased preictal synchronisation by calculation of coherence and pole trajectories. The method has been tested on simulated EEG data and on real EEG data from patients with primary generalised epilepsy. Prediction times of 1-6 s have been found in several seizures from five patients.

No abstract