A robust approach for sleep stage classification using horizontal visibility graphs and a single EEG channel

Detecting sleep stages is an essential part of sleep research; inaccuracies in the electroencephalography (EEG) signals will lead to many issues, including inaccurate disease detection, errors in medication description, and incorrect interpretations of a patient's EEG recordings. The proposed study aims to build a novel method for EEG sleep stages classification depending on a Horizontal Visibility Graph (HVG). The majority of current sleep stage classification approaches depend on frequency and time features. This study suggested a method using the HVG and discriminated features for sleep stages identification utilizing a single EEG channel. Firstly, the EEG signals are transformed to HVG, then extract five discriminated features (edgeCt, PathLength, diameter, degrees(1 to 10), and clustering (1 to 10). All features extracted from a single EEG channel (Pz-Oz) are passed to the Least Squares Support Vector Machine (LS-SVM) classifier. Based on the box plot, the convenient features are selected for each two-class pair to obtain the best classification accuracy. The 10-fold approach was used to assess the performance of the suggested model. Cross-validation of the accuracy of 98.10% for awake and Rapid Eye Movement (REM) two-class pairs.