Based on the nonlinear time series prediction，a feature extraction method for epileptic EEG signals using nonlinear prediction is proposed to automatically detect the epileptic EEG from EEG recordings. To reconstruct the phase space，the approach of determining the embedding dimension based on nonlinear predictability is used to determine the embedding dimension of the EEG signals. The experimental results show that the feature extracted with the method based on nonlinear prediction could clearly distinguish the epileptic EEG from the normal EEG，and the proposed nonlinear feature extraction method is fit for the small set time series and is stable to noise.