Modeling Perceived Sleep Quality Through Objective and Subjective Data

People have different perceptions of what is a good sleep, and assessments of sleep quality are heavily influenced by personal habits. Therefore, there is a considerable gap between what is objectively measured as good sleep with respect to what is, instead, subjectively perceived as such. Polysomnography (PSG) has long been the gold standard for sleep analysis, providing detailed insights into sleep architecture. However, its cumbersome setup limits the ability to capture longitudinal patterns, required to understand what influences perceived sleep quality. Wearable sleep-tracking devices hold the potential to bridge this gap by enabling continuous monitoring over extended periods, gradually adapting to individual personal variations. To fully leverage this capability, it is essential to incorporate predictive features into machine learning (ML) models that account for past personalized trends, including both physiological, subjective, and contextual information. In this study, we monitored a cohort of 23 workers for 30 days, collecting data from wearable devices and daily questionnaires. The best ML model, evaluated on a nested leave-one-subject-out cross-validation, achieved a median ROC AUC of 0.85 employing both absolute and trend-based features. Our preliminary research showcases that the performance of ML models in predicting subjective sleep quality can be significantly enhanced by taking into account personalized health trajectories over time, allowing for more effective longitudinal monitoring.