Time series feature evaluation in discriminating preictal EEG states

Abstract

Statistical discrimination of states in the preictal EEG is attempted using a large number of measures from linear and nonlinear time series analysis. The measures are organized in two categories: correlation measures, such as autocorrelation and mutual information at specific lags and new measures derived from oscillations of the EEG time series, such as mean oscillation peak and mean oscillation period. All measures are computed on successive segments of multichannel EEG windows selected from early, intermediate and late preictal states from four epochs. Hypothesis tests applied for each channel and epoch showed good discrimination of the preictal states and allowed for the selection of optimal measures. These optimal measures, together with other standard measures (skewness, kurtosis, largest Lyapunov exponent) formed the feature set for feature-based clustering and the feature-subset selection procedure showed that the best preictal state classification was obtained with the same optimal features. © Springer-Verlag Berlin Heidelberg 2006.