Ensemble Feature Selection Method Using Similarity Measurement for EEG-Based Automatic Sleep Staging

Sleep staging is the first step in the diagnosis of sleep disorders. In recent years, the automatic sleep staging algorithm based on single channel electroencephalogram (EEG) signal has received extensive attention, but having the problems of insufficient information, poor interpretability and low accuracy. To address above problems, this paper proposes an ensemble feature selection method based on the similarity measurement. Firstly, statistical transformation and the envelope extraction are used to transform the original features obtained by different feature extraction methods in order to obtain candidate features with the desired similarity. Secondly, multiple similarity metrics are used to extract a series of highly interpretable features for sleep staging. Then, two-stage voting approach is used to select the customized features for each subject and the universal features for different subjects. Experiment results show that the proposed method can achieve the classification results with the accuracy of 97.33% when using random forest as the classifier.