Classification of sleep stages using PSG recordings

Polysomnography(PSG) is a test conducted by recording, during overnight sleep, patient's breathing, blood oxygen level and signals read from particular body parts (head, face, heart etc.) for diagnosis and treatment of a variety of disorders related to sleep. Scoring is the first phase required to implement in assessement of recordings. Sleep scoring is an important part of sleep staging, Sleep staging plays a crucial role in diagnosis and treatment of respiratory and sleep disorders. Today, the phase of sleep staging is scored by using polysomnography method. First standard method of scoring sleep recordings was developed by Rechtschaffen and Kales (R&K) in 1968. In this method, the entire polysomnography is divided into 30-second parts called epoch and each epoch is labeled. Each label represents one of the basic stages of sleep. These stages are identified by examining mulitple signals particularly recordigns of Electroencephalogram (EEG), Electrooculogram (EOG) and Electromyogram (EMG). In 2007, The American Academy of Sleep Medicine (AASM) published a scoring manual which provided standardization in sleep recording methods and scoring, in 2012 with the revised version of manual they updated scoring criteria and thanks to this manual sleep stages scoring took its final form and provided the basis of scientific studies. In this study, because of the importance of PSG recordings in sleep scoring, it is aimed at scoring of sleep stages using EEG, EOG and EMG signals. In the laboratory of Necmetttin Erbakan University Meram Faculty of Medicine EEG(C4A1), Chin EMG(CHIN), left eye EOG (LEOG) and right eye EOG (REOG) signals were obtained from 124 patients. These signals were scored by a sleep specialist as Awake, Non-REM-1, Non-REM-2, Non-REM-3, REM and based on these signals scored features have been extracted in time, frequency and non-linear domain and classified. Signals have been classified in five levels as Awake (Wake), Non-REM-1(N1), Non-REM-2(N2), Non-REM-3(N3), REM. Decision Tree (DT), k-Nearest Neighbor Algorithm (kNN), Artificial Neural Network (ANN), Bagged Tree (BT), Support Vector Machine (SVM) algorithms have been used as classification algorithm and feature selection has been made by Relief-F method. Bagged Decision Tree Algorithm has been the most succesful classification system of the three studies which have been carried out. In the last full comprehensive study, 135 features have been prepared for the total dataset and in consequence of classification, Bagged Decision Tree Algorithm had the highest accuracy rate with 84.19%. At the stage of feature selection with Relief-F method, it occured that the highest classification accuracy is Bagged Decison Tree with 14 features and it is observed that it was classified with 95.06% accuracy rate.