Data Sheet 1_Predicting sleep quality with digital biomarkers and artificial neural networks.pdf

IntroductionModern society's increasing stress and irregular lifestyles have led to rising insomnia prevalence, making sleep quality assessment crucial for health management. This study investigates the relationship between heart rate variability (HRV) collected from wearable devices and sleep quality, specifically focusing on wake-after-sleep-onset (WASO) as a critical marker of sleep fragmentation. We aimed to develop predictive models for next-day sleep quality using continuous digital biomarkers.
MethodsWe conducted two experiments (winter and summer 2023) with 82 participants who wore Samsung Galaxy Watch Active 2 devices during wakefulness. Biometric data including HRV signals, daily step counts, and physiological indicators were collected alongside subjective questionnaire responses (PHQ-9, GAD-7, ISI, KNHANES, WHOQOL-BREF) and daily sleep logs. We analyzed seven days of preceding data to predict next-day WASO using various machine learning approaches including ARIMA, Random Forest, XGBoost, GRU, TCN, Transformers, and LSTM models.
ResultsAmong HRV features, the low-frequency to high-frequency (LF/HF) ratio emerged as the strongest correlate with WASO, showing statistically significant differences between groups (Lower LF/HF: 7.5±2.0 min vs. Higher LF/HF: 14.9±3.0 min, p=0.012). LSTM demonstrated superior predictive performance with 90.4% accuracy, 91.3% precision, and 89.9% recall for binary WASO classification. LIME analysis confirmed that LF/HF ratio, along with ISI and WHOQOL-BREF scores, were the most influential features for model predictions.
DiscussionThis work introduces a novel approach for managing sleep health through continuous HRV monitoring and predictive modeling using wearable devices. The findings highlight the potential of the LF/HF ratio as a digital biomarker for sleep quality prediction, offering promise for personalized, data-driven healthcare interventions. The superior performance of deep learning methods underscores the value of temporal pattern recognition in sleep quality assessment, paving the way for proactive sleep health management in everyday life.