Comparison of different clustering algorithms applied to nonliner features for sleep stages discrimination

This study investigates the structure of human sleep at the level of sleep stages. The aim of this study is to demonstrate the effectiveness of nonlinear dynamic features in combination with common clustering algorithms to automate sleep staging. Each 30-second epoch of an overnight sleep EEG were represented by a feature vector of Lyapunov exponent, correlation dimension and entropy. Extracted feature vectors were then presented to one of the clustering algorithm including hierarchical clustering algorithm, expectation maximization, Gaussian mixture and fuzzy c-means clustering. A post processing was also applied to enhance the clustering result. The best clustering over nonlinear measures was achieved by fuzzy c-mean which yields an overall performance of 81% compared to manual scoring of 5 subjects.