A machine learning approach to classify vigilance states in rats

Identifying mammalian vigilance states has recently become an important topic in biological science research. The vigilance states are usually categorized in at least three states, including slow wave sleep (SWS), rapid eye movement sleep (REM), and awakening. To identify different vigilance states, even a well-trained expert must spend a lot of time analyzing a mass of physiological recording data. This study proposes an automatic vigilance stages classification method for analyzing EEG signals in rats. The EEG signals were transferred by fast Fourier transform before extracting features. These extracted features were then used as training patterns to construct the proposed classification system. The proposed classification system contains two functional units. The first unit is principle component analysis (PCA) method, which is used to project the high dimensional features into the lower dimensional subspace. The second unit is the k-nearest neighbor (k-NN) method, which identifies the physiological state in each EEG signal epoch. Based on the results of analyzing 810 epochs of EEG signal, the proposed classification method achieves satisfactory classification accuracy for vigilance states. Based on machine-learning algorithms, the classifier learns to approach the configuration that best fits the categorization task. Therefore, additional training in searching best parameters and thresholds can be avoided. Moreover, the PCA algorithm projects data instances into a 3-D space, making it possible to visualize state-changing dynamics. Experimental results show that the proposed machine-learning based classifier performs better than conventional vigilance state classification algorithms. The results also suggest that it is possible to identify the vigilance states with only EEG signals using the proposed pattern recognition technique.