Automated Multi-Stage Sleep Classification Using Wrist-Worn Photoplethysmography Signal

Sleep plays a vital role in human life. Classifying sleep stages is important for measuring sleep quality as well as detecting any sleep disorders. Traditional sleep staging requires a complex medical setup; several sensors are used to record polysomnography signals, which is both expensive to collect and challenging for participants for long term monitoring, as it may disrupt their sleep. In this study, we classified different sleep stages, including Wake vs. Sleep, Wake vs. Non-Rapid Eye Movement (NREM) vs Rapid Eye Movement (REM), and Wake vs. Light Sleep vs. Deep Sleep vs. REM. We utilized data from 10 participants in the DREAMT dataset, extracting 48 different features from the BVP (PPG) signal across five feature categories: waveform morphology, respiratory modulation, heart rate variability, statistical and energy measures, and complexity and fractal dynamics. We classified the data into two-stage, three-stage, and four-stage categories using the Random Forest Algorithm, k-nearest neighbors algorithm, and Extreme Gradient Boosting (XGBoost) algorithm. Among these classifiers, the XGBoost classifier outperformed. The classification results for the two-class, three-class, and four-class problems using the XGBoost algorithm were 86.35%, 78.20%, and 75.20%, respectively. These findings demonstrate that wearable PPG-based sleep staging can provide a reliable, low-cost, and non-intrusive alternative to traditional polysomnography for sleep monitoring.