Temporal-context learning for subject-independent sleep staging with non-contact multi-frequency radar

Contactless radar may support unobtrusive sleep monitoring, but reliable five-stage evaluation requires participant-isolated testing and a clear assessment of the contribution of model components. The analysis included all 32 overnight recordings (33 964 labelled 30 s epochs) in the PhysioNet Sleep Bioradiolocation Database. Eight documented I/Q channel pairs spanning 3.6-4.0 GHz underwent complex clutter suppression, phase demodulation, fixed 0.05-0.70 Hz zero-phase filtering and resampling to 10 Hz. A 266 342-parameter composite network combined a shared one-dimensional encoder, cross-frequency contrastive initialization, quality-conditioned frequency fusion and a bidirectional gated recurrent unit over nine epochs. Eight deterministic folds each used 24 training, four validation and four held-out test participants, so every participant was tested once. Primary outcomes were participant-level five-stage macro-F1, balanced accuracy and Cohen's kappa. The full network achieved accuracy 0.434, macro-F1 0.376 (95% confidence interval 0.344-0.409), balanced accuracy 0.432 (0.405-0.457) and kappa 0.271 (0.235-0.308). Removing temporal context reduced macro-F1 by 0.072 (0.050-0.095), whereas neither contrastive initialization nor learned fusion showed an incremental held-out benefit. An independently trained early-fusion CNN-BiGRU achieved macro-F1 0.296. The participant-isolated evaluation identifies temporal context as the main contributor among the assessed components and provides a reproducible reference for future radar-based sleep-staging studies.