A test of sleep staging systems in the unrestrained chimpanzee

Abstract

Discrimination and cluster analysis were performed on two sets of sleep data from two chimpanzees. The method of quantification for all periods was spectral analysis of 1 c/sec resolution. One data set (training data set) consisted of selected 10 sec epochs of sleeping EEG which conservatively or liberally represent archetypes of the light-medium-deep sleep staging system. The other data set (testing data) consisted of approximately 100 samples of 10 sec epochs taken systematically throughout the night of sleep. These data are unedited (except for gross artifact) and do not reflect any particular sleep staging system, i.e., they are not preclustered (NPC). Discrimination studies proceeded by the construction of discriminant functions on the training (archetype) set with classification of NPC test data by the discriminant functions so determined. This design was extended to study the handling of transition cases by allowing two types of requirements for archetype admission. Conservative standards minimize within-group variation, but poorly represent the NPC sample. Liberal standards better represented the range of variation in the NPC sample, but discrimination error rates in NPC were equally high with either degree of training set preclustering. Many misclassified test epochs were found to not closely resemble any of the archetype groups. These findings lead us to conclude that neither realization of the LMD system correspond to discriminable clusters in the NPC test data. The cluster analysis section of the study was aimed at: (1) identifying ‘natural’ groupings in the unselected data; (2) taxonomy generation for the archetype data; (3) examination of the archetypes as an internal check on the method's ability to find groups in preclustered data; and (4) a partial investigation of parameters important to cluster analysis as a formal technique ( e.g., the co-ordinate system and its relationship to the distance metric, normalization and scaling, number of clusters, etc.) The results confirmed the method's ability to locate groups in the conservative archetype set. It also demonstrated the relative lack of compactness and localization in the liberalized archetype set as compared to the conservative archetype set. Cluster analysis of the NPC periods suggested that the underlying structure of this data is best described as two large, diffuse clusters denoted as desynchronous (Desyn.) and synchronous (Syn.) (on the basis of the gross appearance of the EEG) with medium cases being distributed within both of these two aggregates. The most compact groups within these two diffuse sets are stage deep for the Syn. aggregate and stage REM in the Desyn. aggregate. Cluster analysis of the liberalized archetypes combined with the systematic samples demonstrated that both archetypes and NPC periods occupy all available clusters. However, there is no dramatic concordance and compactness in the combined results except for deep and REM. The Desyn.-Syn. picture persists with percentual spectra as the data base. We conclude that, within the framework of spectral description from one EEG channel, only these two diffuse clusters can be reliably identified in non-preclustered epochs.