||The sleep staging is the basic core technology in the study of sleep medicine. Via this technology, we can obtain the steep structure and use the results to help us to study different sleep related problems. The architecture of the sleep stages is often altered by various sleep disorders and physical or mental illnesses. Among all the sleep stages, the rapid eye movement (REM) sleep is particularly sensitive to these factors. In fact, many research results have demonstrated the relevance between REM sleep and the functions of cognition and learning. Therefore, if we can accurately detect REM sleep, by comparing REM parameters in different health conditions, we can generate valuable information to assist the medical specialists to evaluate the changes of the health conditions and adaptively adjust the medical treatment procedures. |
The purpose of this thesis is to use a method of machine learning to perform automatic REM sleep detection. Compared to previous studies, this work has several unique features. To detect REM sleep, most of the previous studies needed to extract features from multiple channels of physiological signals. In comparison, this study only require a single channel of EEG signal and thus provide a convenient foundation for home care applications. With the support of the sleep centers, more than five hundred overnight sleep records have been used to develop the proposed approach. The size of the database is considerably larger than those used by the previous work.
By using a deep learning method, we achieve a detection accuracy of 92.26 with a corresponding kappa coefficient of 0.696. By developing a neighboring rule and a compensation rule, this study further improved the results to a kappa coeﬃcient at 0.848, an accuracy of 95.91%, a sensitivity of 86.60%, and a positive predictive value of 87.71%. We have also tried to apply the proposed approach to different subgroups of patients. We have also tested the generalization capability of the proposed approach by using the REM detection method, which was developed by using data of one sleep data, to sleep data of another sleep data. The test results show that the performances of the proposed method is very similar in dealing with data from different sleep centers.