Open Access Dissertation
Date of Award
Electrical and Computer Engineering
Robert F. Erlandson
Sleep is not a single homogeneous state; it consists of several identifiable states. The states are identified by the electroencephalogram (EEG), electrooculogram, and the electromyogram. The most common classification method consists of six classes labeled: wake, stage 1 through 4, and stage REM. A fixed length of data, usually 20 or 30 second epochs, are analyzed and placed in one of the six categories. An all-nigh sleep recording consists of 900 to 1500 such epochs and the task of scoring a record is tedious. Additionally, much of the detail in the EEG is lost due to the size of the epochs and the fixed number of classes. The maximum entropy sleep model is based only on the information contained in the EEG. The EEG is considered as a stochastic process and segments of stationary EEG are identified based on the local properties of the signal. The local properties are estimated by modeling a portion of the EEG short enough to be considered time invariant. These small sections of EEG are compared to adjacent sections using principles from large deviation theory to determine boundaries of stationarity. The stationary segments are classified based on their closeness to a class descriptor. The class descriptor are created by modeling very stable sections of the EEG under analysis. If a stationary segment is not close enough to any existing class a new one is formed, thus each EEG analyzed will have a unique set of classes. The sleep model identifies a microstructure in the sleep record with greater resolution than traditional methods. A macrostructure model based on the microstructure is also presented. The macrostructure model is closer the level of traditional sleep staging. The model was applied to five all-night sleep recordings. The mean length of a stationary segment was found to be 1.5 seconds with only 7 .l% of the segments longer than five seconds. The number of classes for a record was in the range 35-46. A statistical comparison of the microstructure classification to the traditional sleep staging was performed using the chi-squared test statistic. For all record it was found that the probability of the two classification systems being independent was zero.
Clark, Russell James, "An information-theoretical approach to the adaptive segmentation and classification of the sleeping electroencephalogram" (1998). Wayne State University Dissertations. 1762.