An improved K-means clustering algorithm for sleep stages classification

Sleep stage scoring is used to evaluate one´s overnight sleep process, which is important for clinical diagnosis. However, the visual inspection of sleep data is a laborious task and the scoring results may be subjective to different clinicians. The purpose of this paper is to develop an automatic sleep stage classification algorithm to reduce the artificial workload. The overnight sleep data are represented by the extracted features from time domain and frequency domain of EEG. An improved k-means clustering algorithm is proposed to classify overnight sleep data into five stages including awake (W), NREM (Non-Rapid Eye Movement) stage 1 (S1), NREM stage 2 (S2), slow-wave sleep (SS) and REM (Rapid Eye Movement). In the improved k-means clustering algorithm, the points with dense surrounding are selected as the original centers by using the concept of density for reference. Additionally, the cluster centers are updated according to `Three-Sigma Rule´ during the iteration. The determination of cluster center selection was developed which can be adaptive to the actual cases and abate the singular points effect. The obtained results showed that the accuracy of proposed algorithm was satisfied; especially it can distinguish W, SS and REM effectively. Furthermore, the improved k-means algorithm had less number of misclassification and higher accuracy than the original algorithm.