Article Preview
Top1. Introduction
Sleep stage classification, otherwise termed sleep scoring or identification, is crucial for treating sleep-related disorders (Wulff, Gatti, Wettstein, & Foster, 2010). Many people with sleep issues are at risk of developing other health problems such as obesity, diabetes, and neuropsychiatric disorders (Kammerer, Mehl, Ludwig, & Lincoln, 2021; Kim, Kang, Choe, & Yoon, 2021; Lu et al., 2021). Polysomnography (PSG) is a medically proven gold standard used to diagnose common sleep disorders (Thorpy, 2017). PSG acquires biosignals from the body like brain signals (electroencephalogram, EEG), movement of the eye (electrooculogram, EOG), heartbeat (electrocardiogram, ECG), and jaw movement, or limb muscular activity (electromyogram, EMG). The trained sleep scoring experts examine the recorded signals and assign a sleep stage to each 30-second PSG data termed as an epoch. The sleep technicians either follow the guidelines of Rechtschaffen and Kales (Allan Hobson, 1969) or the American Academy of Sleep Medicine (AASM) (Richard B. Berry, MD; Rita Brooks, MEd, RST, RPSGT; Charlene E. Gamaldo & Susan M. Harding, MD; Robin M. Lloyd, 2016). In the manual approach, sleep experts have to visually evaluate epoch by epoch and label sleep stages to build a hypnogram indicating corresponding sleep stages. Sleep stages are defined as wake-state (W), rapid eye movement (REM), and non-REM (NREM). According to AASM, NREM comprises three stages N1, N2, and SWS (slow-wave sleep). Manual sleep scoring has drawbacks like time-consuming, labour-intensive, the requirement of highly trained sleep technicians, inter-rater variability and occasionally subjective (Cesari et al., 2021; Stepnowsky, Levendowski, Popovic, Ayappa, & Rapoport, 2013). Furthermore, overnight PSG studies have limitations such as expensive, big waiting lists in the clinics, unfamiliar sleeping environments, restricted privacy, skin irritation due to the adhesive from the electrode and multiple sensors connected to the person may obstruct sleep, lowering the recording accuracy (Zhang et al., 2021). Developing a robust and convenient automated sleep stage classification system could be highly beneficial to overcome the limitations mentioned above. The role of EEG signals is significant in identifying specific sleep stages among all the distinct signals recorded by PSG, either in manual or automated scoring methods (Kwon, Kim, & Yeo, 2021). Recent technological improvements play a significant role in designing and developing reliable in-home based automatic sleep stage classification (ASSC) systems (Kwon et al., 2021). This study proposes and uses a single-channel EEG to classify different sleep stages to develop a practical in-home sleep system, as prior studies have shown (Ghimatgar, Kazemi, Helfroush, & Aarabi, 2019a; Hassan & Bhuiyan, 2017a). Thus far, single-channel EEG based sleep classification approaches, particularly machine learning algorithms, have been widely researched.
Typical machine learning algorithms follow a traditional procedure of preprocessing, feature extraction, and selection of features before passing the data to the classifier. The major ASSC works in the literature rely on handcrafted features from the domain such as time, frequency and time-frequency (Boostani, Karimzadeh, & Nami, 2017a). In the time domain, major features obtained includes statistical features (mean, variance, standard deviation, skewness, kurtosis, etc.), Hjorth parameters, threshold and zero-crossing rate. In the frequency domain, frequently extracted features comprise spectral estimation, parametric methods (autoregressive, moving-average, autoregressive moving-average), non-parametric approaches (power spectral-density) and higher-order spectra. The most important features extracted from the time-frequency domain includes entropy and complexity based (Renyi’s, Tsallis, permutation, Lempel–Ziv, multi-scale, approximate etc.), and fractal-based (correlation dimension, Lyapunov exponent, Hurst exponent, Petrosian, Higuchi etc.) (Boostani et al., 2017a), (Zhao, 2019).