| The study of sleep staging is of important clinical significance for detection, prevention and treatment of sleep disorders. The traditional manual interpretation based methods are very laborious and lack of consistent and objective criteria. Moreover, due to the existence of artifacts and non-linearity of sleep EEG signals, the existing signal processing methods cannot produce satisfactory results in the problem of sleep staging. In order to solve the difficulty of feature extraction and the inefficiency of the traditional methods, in this dissertation we employ approximate entropy (AE) and sample entropy (SE) as features to differentiate the different characteristics during various sleep stages. Because of the high time complexity of approximate entropy and sample entropy, the method proposed here is only applicable to off-line cases. To test the performance of the new method, real clinical data collected from subjects during sleep study are utilized in this study. The main contents in this dissertation include:1. Preprocessing of the sleep EEG signals that are time-varying, non-stationary, weak in amplitude and full of noise and artifacts. The purpose of preprocessing is multifold. Firstly, it is mandatory to filter out the 50Hz power-line interference. This is accomplished by applying an infinite impulse response (IIR) filter to the sleep EEG signals. The output of this IIR filter is then fed into a module of wavelet packet decomposition so as to cancel the possible EMG interference. The cleaner sleep EEG signals after such kind of preprocessing are then ready for further sleep staging.2. Sleep staging based on two indices, namely AE and SE, which are defined in the nonlinear dynamics theory. This part of study consists of the definition and algorithmic design of AE and SE along with their application in processing the cleaner sleep EEG signals resulted from the preprocessing procedure described above. Experiment results have shown that both AE and SE can be used as discriminatory features for the sake of sleep staging. A comparative investigation suggests that SE outperforms AE in terms of amplitude increase—the amplitude increase of SE is 20% to 35% higher than that of AE. Moreover, SE is more consistent than AE. After extracting AE and SE from the sleep signals, some support vector machines (SVMs) are then designed to classify the different stages of the sleep signals. To reduce the number of SVMs and hence the time complexity of the whole algorithms, we build a binary classifier tree (BCT) whose nodes are all SVMs. Experiment results show that the BCT developed can distinguish between awaking and other sleep stages with an accuracy of 100%, demonstrating the feasibility of using SVM for the purpose of sleep staging. A combination of SE and AE can obtain accuracy as high as 80%, better than using either SE or AE alone in classifying other difference sleep stages. |
