A New Investigation of Automatic Sleep Stage Detection using Decision-Tree-Based Support Vector Machine and Spectral Features Extraction of ECG Signal

抄録

This research is proposed to investigate an easy, fast, and effective method for automatic sleep stage detection using spectral features extraction from electrocardiography (ECG) signal alone. Sleep stage detection is the gold standard for sleep analysis. A sleep physician may suspect a treatment and diagnosis of sleep diseases through sleep stage detection. Polysomnography (PSG) method that generally used for detecting sleep stage. This method is intrusive and difficult to be implemented for in-home and portable systems. Moreover, it is involved a lot of effort, time and cost. The automatic sleep stage detection that uses only ECG signal is expected can solve these problems. The proposed method was tested and evaluated on 51 subjects that consist of 16 healthy subjects, 9 patients with insomnia, 4 patients with sleep-disordered breathing, and 22 patients with REM behavior disorder. In this research also tried to apply a minimal feature for automatic sleep stage detection. The two features applied were the normalized Low Frequency, LF (n.u.) band power and normalized High Frequency, HF (n.u.) band power that obtained from spectral features extraction. These features were then used as inputs for sleep stage classification. Mostly commonly used learning classifiers is implemented to classify sleep stage, namely KNN, NN, DT, SVM, and proposed DTB-SVM. The proposed method using DTB-SVM and spectral features extraction of ECG achieved an average classification specificity, sensitivity, and overall accuracy of 98.31%, 91.84%, 95.06%, respectively. The proposed method is able to obtain all sleep stage condition on patients and non-patients subjects. However, it is feasible to implement in-home and portable system of automatic sleep stage detection instead of using a multichannel signal.