EEG Features as Biomarkers for Discrimination of Preictal States

The aim of this study is the selection of the most relevant features of electroencephalograms (EEG) for classification and clustering of preictal states. First, a sum of 312 time series features were computed on consecutive segments of preictal EEG (simple statistical measures, linear and nonlinear measures), where some of them regard different method specific parameters. The efficiency of three methods for feature selection was assessed, i.e., the Forward Sequential Selection (FSS), Support Vector Machines with Recursive Feature Elimination (SVM-RFE) and a MI filter. The classification was applied first to 1000 realizations of simulated data from the Mackey–Glass system at different high dimensional chaotic regimes, and next to 12 scalp early and late preictal EEG recordings of different epileptic patients (about 3 h and half an hour before the seizure onset, respectively). The optimal feature subsets selected by the three feature selection strategies for the same classification problems were found very often to have common features. Based on these feature subsets, classification with k-means partitioning as well as SVM was assessed on test sets of EEG from the same recordings. Feature subsets for each channel and episode or only episode did not classify on the test set as well as a global feature subset of a sufficiently large number of the most frequent features over all channels and episodes. We concluded that a global feature subset of 16 most frequent features can play the role of a biomarker and distinguish early and late preictal states.