Improved Sleep Stage Tagging on Wearables via Knowledge Distillation

Sleep disorders present a major global health challenge, requiring precise and efficient diagnostic approaches. Polysomnography (PSG) remains the gold standard by capturing multiple physiological signals during overnight monitoring. However, PSG is both resource-intensive and burdensome for patients. A key limitation is the so-called first-night effect, i.e. the altered sleep architecture due to the unfamiliar laboratory environment and sensor setup. This can lead to artificially reduced sleep efficiency and diagnostic distortion. Wearable-based sleep monitoring offers a promising alternative, enabling longitudinal assessment and reducing the first-night effect by using fewer sensors and enabling home-based diagnostics. Recent advancements in wearable technologies have improved accuracy in estimating sleep parameters when compared to PSG. This study investigates knowledge distillation (KD) as a strategy to transfer knowledge from an existing high-performing PSG-based teacher model to a student model operating on wearable-like input data. Both models use the UTime architecture which has been shown to perform well across different sleep laboratories and are trained on clinical PSG data from a sleep laboratory in Germany. Our KD framework combines soft target supervision and feature matching to guide the student’s learning. The results show that KD consistently improves the performance of the student in various sensor configurations and model sizes, indicating its potential for accurate sleep staging in real-world wearable applications.