We apply a nonlinear prediction algorithm to investigate the presence of no
nlinear structure in electroencephalogram (EEG) recordings. The EEG signal
could be modeled as a realization of a nonlinear model plus a residual nois
e (uncorrelated Gaussian noise). Using linear and nonlinear models we analy
ze the statistical nature of these residual noises in the case of epileptic
patients and normal subjects. We found that the residual noise presents Ga
ussian distribution for epileptic patients if a nonlinear model is used whe
reas in the case of normal subjects the residual noise will exhibit a Gauss
ian distribution only if a linear model (autoregressive) is used. These res
ults provide another evidence of the nonlinear character of the epileptic s
eizure recordings, while the normal EEG seems to be better described as lin
early correlated noise. (C) 2001 Elsevier Science B.V. All rights reserved.