An Efficient Sleep Scoring Method using Visibility Graph and Temporal Features of Single-Channel EEG

This work proposes a method utilizing the fusion of graph-based and temporal features for sleep stage identification. EEG epochs are transformed into visibility graphs from which mean degrees and degree distributions are obtained. In addition, autoregressive model parameters, Higuchi fractal dimension, multi-scale entropy, and Hjorth’s parameters are calculated. All these features extracted from a single EEG channel (Pz-Oz) are fed to an ensemble classifier called random undersampling with boosting technique. Two different approaches i.e. 10-fold crossvalidation and 50%-holdout are utilized to evaluate the performance of the model. Cross-validation accuracies of 91.0% and 97.3%, and kappa coefficients of 0.82 and 0.94 are achieved for 6- and 2-state classifications, respectively, which are higher than those of existing studies.Clinical relevance— Automatic and reliable sleep stage classification can reduce the burden of sleep experts in analyzing overnight sleep data (~ 8 hours). It can also assist them to find specific traits of interest such as spindle density, by providing annotated sleep data (hypnogram), thereby eliminating the need for tedious and expensive manual scoring. An accurate 2-state (wake/sleep) classification is also crucial for the patients with disorders of consciousness, where stimulation during wake state is considered more effective than that in sleep state.