Contactless sleep staging with Radar: a transfer learning approach

Accurate sleep monitoring is essential to assess sleep quality and diagnose sleep disorders. Although sleep laboratories provide precise assessments, they are expensive, labor-intensive, and unsuitable for long-term or large-scale monitoring. Radar-based sensing offers a fully contactless alternative, enabling unobtrusive real-world sleep monitoring. However, the lack of large, labeled datasets has limited the development of robust sleep stage classification models. We address this with transfer learning to improve classification accuracy and generalization to unseen participants within the radar cohort. An LSTM model was pretrained on movement, HRV, and respiratory features from the MESA Sleep dataset (>1,100 participants) and fine-tuned using radar data from 44 synchronized polysomnography recordings. Transfer learning increased the Matthews Correlation Coefficient from 0.25 to 0.47 (five-class staging), particularly for Wake, N3, and REM sleep. Future work should explore domain-adaptation across modalities and cohorts. Our results highlight the potential of radar-based sleep analysis for scalable, contactless long-term sleep monitoring.