基于正态逆高斯和特征贡献度的睡眠分期实验研究

An experimental framework based on normal inverse Gaussian and feature contribution was proposed for automatic classification of sleep stages. Features were extracted from the sleep EEG (electroencephalo-graph) signals. The signals were decomposed by tunable Q-factor wavelet transform (TQWT). The normal inverse Gaussian parameters were extracted from the TQWT sub-bands. The important features were selected and ranked according to the contribution degree based on the SVM model; according as the selected features of high contribution, the results of different classifiers were compared afterwards. A multi-classifier based automatic sleep staging algorithm was then designed. Results show that, the accuracy of sleep staging can reach 89.88% according to the validation on sleep-EDF dataset from PhysioBank. Compared with the single classifiers, the accuracy of staging can be improved greatly. Therefore, the proposed method is of great value for the clinical diagnosis and researches of sleep disorders.