PPG-Based Sleep Stage Classification Using Pulse Wave Feature Fusion and Explainable AI

Abstract

Sleep monitoring plays a crucial role in understanding and managing various health conditions, including sleep disorders, cardiovascular diseases, and mental health. Traditional sleep monitoring methods rely on Electroencephalography (EEG) and Polysomnography (PSG) in clinical settings. However, these methods are expensive, difficult to administer, and unsuitable for home-based monitoring. In recent years, photoplethysmogram (PPG) has emerged as a promising noninvasive technology that is widely used in wearable devices and holds great potential for sleep assessment. Yet, most current sleep monitoring methods rely on deep learning models, which are inherently "black-box" and challenging in the clinical decision-making process. In this paper, we propose an explainable random forest model for sleep stage classification using pulse wave feature fusion. Our method employs statistical, temporal, and nonlinear dynamical features extracted from the PPG pulse wave associated with sleep patterns. Additionally, we investigate the digital biomarkers of sleep and PPG using SHAP (SHapley Additive exPlanations) methods to enhance interpretability. The proposed approach demonstrates competitive performance, achieving an overall accuracy of 82.56% in two-stage (sleep and wake) classification, 77.79% in three-stage (wake, NREM, REM) classification, and 69.20% in four-stage (wake, light sleep, deep sleep, REM) classification. The results highlight the potential of PPG-based wearable devices in sleep monitoring, offering a feasible solution for home-based assessments with clinical applicability.