Passive and continuous assessment of subjective sleep quality using smartwatches

Sleep is a fundamental physiological process that directly impacts nearly every aspect of human health. Poor sleep quality has been linked to cardiovascular disease, stress, anxiety, depression, inflammation, impaired cognitive performance, and even increased mortality. When assessing sleep quality, it is important to capture two fundamentally different yet complementary dimensions: objective sleep quality and subjective sleep quality, which are both independent predictors of adverse outcomes. While advances in wearable technology now allow continuous monitoring of objective physiological characteristics of sleep, subjective sleep quality still relies exclusively on self-reported questionnaires, which inherently require active user participation and are limited to specific points in time. In this study, we present a machine learning framework for the passive and continuous assessment of subjective sleep quality, leveraging physiological signals collected from commercial smartwatches, such as sleep data, physical activity, heart rate, and stress inferred from heart rate variability, alongside individual characteristics, including sociodemographic factors, personality traits, and health-related attributes. Our machine learning model was trained using data from the PerMed prospective observational study, which includes 4,794 participants, 759,116 self-reported daily questionnaires and 1,765,264 days of smartwatch recordings collected over a three-year period. We show that the model achieves an area under the receiver operating characteristic curve of 0.88 in distinguishing between extreme subjective sleep quality ratings. Our passive and continuous approach to assessing subjective sleep quality represents a conceptual shift in how subjective sleep quality is measured, with the potential to enable timely, personalized interventions that support not only sleep-related conditions but also broader aspects of overall health and well-being. How well people feel they have slept can affect their health and well-being. However, understanding this usually requires people to regularly report their sleep quality themselves. We investigated whether this information could instead be estimated automatically using commercial smartwatches. We analyzed three years of data from 4,794 participants, combining smartwatch measurements of sleep, physical activity, heart rate, and stress with personal characteristics such as age, health, and personality. We developed a machine learning model that could distinguish between nights that participants rated as very good or very poor. Combining smartwatch measurements with personal characteristics performed better than using either type of information alone. These findings suggest that smartwatches may eventually help track how people experience their sleep continuously, with less need for repeated questionnaires. Hausner et al. develop a machine learning framework combining commercial smartwatch data with individual characteristics to assess subjective sleep quality. The model achieves high performance in distinguishing extreme levels of self-reported sleep quality, opening the possibility of passive and continuous assessment of subjective sleep quality.