Study Of Epileptic Seizure Prediction Algorithms Based On EEG Signal

Epilepsy is a common brain disease state. There are plenty of patients in the world suffering from epilepsy and some of them can’t be cured. Epilepitic seizure prediction can help to improve these people’s life. The study of seizure prediction is focus on the detection of epilepitic seizure and prediction with single method presently. Basing on this situation, we propose a kind of algorithm which combined signal preprocessing and pattern recognition. The performance of this algorithm is verified, such as accuracy, sensitivity, false alarm rate and forecast interval. At last, simulated online dection is applied to test the algorithm.Basing on the former researchers’ study and the dataset we used in this paper, the algorithm consists of two parts: signal preprocessing and pattern recognition. In the signal preprocessing part, EEG’s energy feature is extracted by wavelet transformation and power spectral. In the pattern recognition part, support vector machine(SVM) is applied to separate the data. The improvement of pattern recognition’s accuracy is expected in epilepitic seizure prediction by the combination of two partsThe research result shows that the energy of frequency band 0.5~8 Hz would rise 2000 seconds before seizure onset by analysing interictal and preictal EEG’s wavelet energy. We used relative wavelet energy and SVM to analyse and test 9 patients’ EEG data, and it shows that the algorithm can predict some patients’ seizure onset except a few of patients’ bad behavior. We replace the wavelet with spectral power in the signal preprocessing part, and use it to extract feature. The predict accuracy is improved by using spectral power and SVM. Comparing to the relative wavelet energy, the result of 6 patients’ test data improved by spectal power.In the end of this paper, we simulate real time seizure prediction and discuss the channels correlation. Spectral power and SVM are applied in simulate real time test and we test a patient’s whole EEG data with the model we build before. It shows that the algorithm alarms 2000 to 5000 seconds before seizures’ onset and we can ignore discontinuous false alarms in test. We select 14 out of 23 channels by correlation analysing. The algorithm is accelerated by using these 14 channels and the performance doesn’t lose. There are 6 patients in the test. By using 14 channels, 2 patients’ accuracy improved and 3 patients’ accuracy keep.