Human Sleep Scoring Based on K-Nearest Neighbors

Abstract

Human sleep is one of the essential indicators that gauge the overall health and well-being. Presently, it iscommon for people to face issues related to sleep. Various biomedical signals including electroencephalogram (EEG),electrooculography (EMG), and electrooculography (EOG) are utilized in the diagnosis and during the treatment ofsleep disorder cases. An automatic classification to diagnose sleep problems can help in the analysis of sleep EEGdata. In this current study, an effort is made to classify the sleep stages from a single EEG channel (C4-A1) basedon K-nearest neighbors (K-NN) with three alternative distance metrics. The Euclidean distance is the most commonlyused distance measure in K-NN, and no prior study of sleep EEG data has inspected the classification performance ofK-NN with various distance measures. Therefore, this study aimed to investigate whether the distance function affectsthe performance of K-NN in the classification of sleep data. Euclidean, Manhattan and Chebyshev distance measureswere individually tested with K-NN classification, and their performances were compared based on accuracy, sensitivity,specificity, F-measure, Kappa statistic and computation time for both Rechtschaffen & Kales and American Academyof Sleep Medicine standard labelings of the sleep stages. The experimental results show that the Manhattan distancefunction with K = 5 was the best choice for classification of the sleep stages, achieving 98.46% and 98.77% correct ratesfor the two labelings with comparatively rapid computations.