EEG-based multivariate statistical analysis of sleep stages.

Abstract

All-night sleep of 5 healthy male subjects was scored on the basis of EEG (paper records), EMG and EOG into the stages of wakefulness, REM sleep and slow sleep stages 1-4 according to Rechtschaffen and Kales. Spectral analysis for 8 EEG channels was then performed and spectral parameters (total power and coherence for certain frequency bands) extracted. Stepwise linear discriminant analysis was applied to these spectral parameters to see how well the polygraphically defined sleep stages were recognized. While the results of each intrasubject analysis were satisfactory (error rates between 10 and 15%), a more detailed analysis revealed that most errors occur between nearby slow sleep stages or between S1 and REM or wakefulness. This is related to the fact that Rechtschaffen and Kales classify into discrete stages an essentially continuous process and this is, to some extent, arbitrary. This interpretation is supported by examples. Cross-classification of each subject on the basis of the remaining 4 increases the error rates drastically. A simple way of avoiding this effect, by standardizing the individual data, is shown to work nicely in this case.