Feature Extraction for Machine Learning-based Sleep Stage Classification Using PPG-Derived Parameters and Skin Temperature

This study explores feature extraction and machine learning methods for automated sleep stage classification using physiological signals, focusing on PPG-derived parameters and skin temperature. We experimented with feature extraction techniques that required different engineering steps, i.e., with or without applying the first derivative and deviation from the mean of raw signals. The statistical values, including mean, standard deviation, maximum, minimum, and median of each 30-second segment of signals, are then extracted and used as input features for our selected classification model, the Extra Trees. The classification performance is evaluated under both subject-dependent and subject-independent scenarios using data from 93 subjects. The data include labeled stages of wake, non-rapid eye movement (NREM) sleep stages N1, N2, and N3, as well as rapid eye movement (REM) sleep. Optimal features were derived from skin temperature measurement along with the normalized timestamp, which is associated with sleep duration. Results show that the subject-dependent experiment yields the highest performance with an accuracy of 0.87 for three-stage classification (wake, NREM:N1/N2/N3, REM) and 0.77 for five-stage classification (wake, N1, N2, N3, REM). This work highlights the potential of artificial intelligence to enable the automatic labeling of sleep stages using non-constrained sensors. The important features related to sleep staging can be investigated, which could contribute to future advances in clinical diagnostics and healthcare applications.