Prediction Of Epileptic Seizures Using The Singular Spectrum Analysis And The Artificial Neural Networks

II The epilepsy is a kind of the brain malfunction, which is caused by the super-synchrony discharge in the cerebral cortical neurons. At present, research proved that seizures can be detected early before the onset of seizures to greatly reduce the harms caused by the seizures and to improve the life quality of the patients. Thus, the forecast method becomes important. So far, the scalp EEG data is very well used in the research. This paper proposes the singular spectrum analysis (SSA) method to investigate the prediction method suitable for some short period of time series. This method is linear, simple and effective in the project.Followings are the details:At first, the pre-ictal EEG recordings are equally segmented. The forecast time, which is defined as the time distance advance the onset of the full seizure, is needed. In ordinary, it is permitted as the maximum values from the experimental experience, here, it is chosen as 100 seconds.Then, the Cao method is used to calculate the phase space reconstruction of the EEG data, which studies the basic nonlinear dynamic system. Here the delay timeτand the embedding dimension m are regarded as the results of the phase space reconstruction. Here,τ= 5 and m = 15 are available in the project.Third, the singular spectrum is calculated based on the phase space reconstructed. It is shown that the healthy people singular spectrum is quickly reduced to a very small value after the initial larger values. This small value range is named as "noise platform." However, for the epilepsy patients, the EEG singular spectrum is slowly declining, unlike that of the healthy people, there is no "noise platform". The cases of the partly focal seizures are also included in the computing. It is concluded that the method can distinguish normal regional and foci zone by the character of the "noise platform".In the end, the back-propagation (BP) neural network model is effectively used as a classifier to sort the health people and the epilepsy patients by the singular spectrum.The method is proved to have the following advantages: As a linear analysis (SSA), it is simpler than the traditional non-linear dynamic analysis methods in the computational complexity. Also, only very short of the pre-ictal EEG data is necessary, for example, four seconds (512 points) here is sufficient, that make the method to be very easy and very practical for the clinical application.