EEG Sleep Pattern Recognition by Cluster Analysis

Abstract

The sleep EEG is a highly individual signal containing an abundance of different patterns; atypical EEG patterns may additionally appear under the influence of psychopharmaca, stress, pain and other stimuli. There-fore a dynamical classification scheme, free of a priori assumptions of universally valid sleep states, is indicated for sleep analysis permit-ting also the detection of unknown patterns. The main features of the EEG are the power and coherence spectra; they form a timevariing vector that moves through the feature space during night. A principal component transformation is performed in order to reduce the dimension of the vector space and to minimize the correlations between the variables. The grouping process of the transformed data is iterative; for an increasing number of groups a nonhierarchica1 cluster analysis in combination with discriminant analysis is performed to evaluate the amount of really separable clusters. The sleep states are defined by the location in the feature space, the covariance matrices, the realisation and transition probabilities. Once the individual sleep states are defined they are used as a basis set for the classification of data from further nights of the same subject. The allocation of a vector to a known sleep state is determined by the smallest Mahalanobis distance on condition that the constraints of homogeneity are not violated.