Classification of sleep stage with period analysis features derived from the EEG.

This paper details our experience with the use of regression analysis and discriminant analyses, combined with a Bayes classifier, to evaluate the usefulness of a series of features derived in real time from a single channel of EEG. These features were 1) the number of zero crossings, 2) the number of zero crossings of the first derivative of the EEG, and 3) the number of zero crossings of the second derivative and five additional "histogram" measures. Each histogram measure represented the number of occurrences of a zero crossing interval falling within an arbitrarily selected range of values. The distributions of several measures conditioned on sleep stage are presented along with night-to-night and subject-to-subject variations in the measures. These details provide a basis for understanding effectiveness and limitations of zero crossing measures and the regression and discriminant analysis techniques. Our results permitted a range of 57-90% accurate classification with just zero crossing measures with the cross-correlation between computer- and hand-scored nightly sleep patterns ranging from 0.63-0.94. We consider these to be good and acceptable accuracies for this type of analysis.