Optimising Sleep Stage Detection Using a Minimal Non-EEG Physiological Signal Set and Deep Learning

Automatic sleep stage classification is essential for enabling non-invasive, at-home monitoring. However, current methods often rely on electroencephalogram (EEG) signals and ad-hoc development approaches that limit reproducibility. We present a reproducible engineering framework for a deep learning model based on the U-Net architecture that classifies sleep into five stages (Wake, N1, N2, N3 and REM) or four (Wake, Light Sleep, Deep Sleep and REM) using only three easily acquired physiological signals: oxygen saturation (SpO), heart rate (HR) and abdominal respiratory effort (AbdRes). In contrast to most previous studies, our model provides sleep stage predictions on a per-second basis, thus overcoming the limitations associated with fixed 30-s epochs. The model was trained on the Sleep Heart Health Study-Visit 2 (SHHS2) dataset and externally validated on the Multi-Ethnic Study of Atherosclerosis (MESA). Optimisation of the model was achieved via Keras Tuner with the Hyperband algorithm. The study achieved weighted F1-scores of 68% (five-stage) and 71% (four-stage) with Cohen's Kappa of 0.61 and 0.67 on SHHS2, with consistent performance on MESA. These results demonstrate strong generalisation and suggest that this lightweight, EEG-free approach offers a practical path towards scalable, clinically relevant sleep monitoring.