EEG-based, neural-net predictive classification of Alzheimer's disease versus control subjects is augmented by non-linear EEG measures.

Attempts to classify Alzheimer's disease (AD) subjects versus controls using spectral-band measures of electroencephalographic (EEG) data typically achieve around 80% success. This study assessed the ability of adding non-linear EEG measures and using a neural-net classification procedure to improve this performance level. The non-linear EEG measures were estimated correlation dimension ("dimensional complexity," or DCx) and saturation (degree of leveling-off of DCx with increasing embedding dimension). In a sample of 39 subjects (14 ADs, 25 controls), it was found that (a) the addition of non-linear EEG measures improved the classification accuracy of the AD/control status of subjects, and (b) a back-percolation neural net predictively classified the subjects much better than the standard linear techniques of multivariate discriminant analysis or nearest-neighbor discriminant analysis.