Automatic Classification of Sleep Stages using EEG Sub-bands based Time-spectral Features

Abstract

Sleep scoring is proved of having major impact on treating various sleep oriented disorders. But achieving this task manually is very time consuming and long process. Hence, an efficient computer based system is required to carry out the epoch based multi-class sleep stages classification. Among all the polysomnography (PSG) signals, Electroencephalogram (EEG) provides valuable information for sleep related analysis by sensing and monitoring the brain functions. Hence in this study, an effective computer-assisted technique is proposed for classifying various sleep stages. Firstly, the input signal is segmented into 30 seconds epochs as per the Rechtschaffen and Kales criteria (1968). From the six EEG sub-bands, five features such as Normalized power, Movement, Mean Absolute Deviation, Inter-quartile range and Fourier Synchrosqueezed transform are extracted. The feature vector is subjected to 10-fold cross-validation for 2-class to 6-class classification. The results are obtained after computing accuracy, sensitivity, specificity and Cohen Kappa's statistics using SVM classifier. The highest accuracy of 98.4%, 95.8%, 94.3%, 93.4% and 92.5% is achieved for 2-class to 6-class classification respectively. Also, subject specific results are computed for 5-class problem for which F1 score is evaluated for each stage. This proposed method offers improved results as compared with other previous studies.