Automatic Seizure Prediction from Scalp EEG with Optimal Feature and Minimum Channels

The aim of this study is to develop an efficient, reliable and automatic epileptic seizure prediction system using scalp EEG measurements with optimal feature and minimum channels. EEG data in interictal and preictal periods from the CHB-MIT dataset are used for seizure prediction. First, the original signals are decomposed into several frequency bands using a digital wavelet transform (DWT). Then, features including standard deviation (S), log of amplitude (L), quartile (Q) and coefficient of variation (CV ) are extracted. Finally, different combinations of feature vectors are fed into classifiers (support vector machine (SVM) and extreme learning machine (ELM)) to classify the above two states (preictal and interictal states). Performance analysis shows that the optimal feature is CV, the optimal sub-band is 16-31 Hz and the optimal EEG channel can be chosen as FP1-F7, T7-P7, FP1-F3, P3-O1 or P7-T7. By comparing the classification results, ELM provides a more robust and higher overall accuracy than SVM, and the best average accuracy of both ELM and SVM can reach as high as 100%.