Nonlinear prediction and complexity of alpha EEG activity

Two prediction techniques were used to investigate the dynamical complexity of the alpha EEG; a nonlinear method using the K-nearest-neighbor local linear (KNNLL) approximation, and one based on global linear autoregressive (AR) modeling. Generally, KNNLL has more ability to predict nonlinearity in a chaotic time series under moderately noisy conditions as demonstrated by using increasingly noisy realizations of the Henon (a low-dimensional chaotic) and Mackey-Glass (a high-dimensional chaotic) maps. However, at higher noise levels KNNLL performs no better than AR prediction. For linear stochastic time series, such as a sine wave with added Gaussian noise, prediction using KNNLL is no better than AR even at very low signal-to-noise ratios. Both prediction techniques were applied to resting EEGs (O2 scalp recording site, 10-20 EEG system) from ten normal adult subjects under eyes-closed and eyes-open conditions. In all recordings tested, KNNLL did not yield a lower root mean squared error (RMSE) than AR prediction. This result more closely resembles that obtained for noisy sine waves as opposed to chaotic time series with added noise. This lends further support to the notion that these EEG signals are linear-stochastic in nature. However, the possibility that some EEG signals, particularly those with high prediction errors produced by a noisy nonlinear system cannot be ruled out in this study.