Eeg Feature Analysis And Feature Extraction

Epilepsy is a common intrinsic neurological disorder. Epileptic seizures often occur suddenly, causing disability and mortality sometimes when the patient loses his awareness. Predicting an impending epileptic seizure has obvious clinical importance. There are many EEG-based signal processing methods used to predict seizures, but better research results are mainly based on intracranial EEG recordings. In comparison with intracranial EEG, the examine method of the scalp EEG has been used in clinical diagnosis universally for its nondestructive. However, the scalp EEG signal is well known to be subject to the noise or artifact contaminations, making the analysis more difficulty. Therefore none of the ideal seizure predicting results on scalp EEG is published up to now. Thus the goal of this thesis is to find appropriate methods to predict the incoming epileptic seizures based on scalp EEG effectively.Multiresolution analysis theory is used to extract appropriate features in this thesis and the recurrent neural network is a classifier. The performances of the synthetical system to predict seizures are as follows:First, the basic theory, the history and the present of seizure prediction based on EEG signals are reviewed, as well as different signal processing methods.Second, the basic theory of wavelet transform is studied and the frequency properties of the orthogonal wavelets are analyzed in detail. According to such performances as the center of the window in time-frequency domain, the width of the window and the concentrating degree of the power in frequency domain, this thesis considers that the wavelet daubechies 5 is more suitable for the application, because it has less frequency overlap between adjacent scales in frequency domain.Third, the scalp EEG signals are decomposed to different rhythm components using multiresolution analysis, and the evolution of the absolute wavelet power spectra of different rhythm in intraictal and preictal stage is explored. Then epileptic EEG and normal EEG are compared by the relative wavelet power spectra to find the abnormities. The result is that the slow wave is enhanced and the a rhythm is weakened hi epileptic EEG signals.Fourth, the architecture properties of the basic Elman network hi the recurrent neural networks is lucubrated, as well as the network training procedure based on the decoupled extended Kalman filter algorithm. According to the characteristic, this thesis proposes an improved method, forcing the feed-forward weights of the context units hi the Elman network to be unit. Theoretical analyses and simulation results show that such improvement can not only decrease the computational complexity and the storage space requirement of the network training algorithm effectively, but also increase the tracking ability to the high order dynamic system of the Elman network.Finally, combined with the wavelet preprocessing, the unproved Elman network isapplied to seizure prediction. Three different signal features are used as network inputs. It can be concluded by comparing the results that for these data sets, extracting the envelope of the power spectra of the a rhythm in wavelet transform domain can improve the performance of RNN effectively.