Fold-Local Validation-Gated Sparse Multi-View Fusion for Apple Watch Sleep Staging

Background: Wrist-worn heart-rate and motion signals may support sleep staging outside the laboratory, but heterogeneous model libraries create substantial scope for optimistic analysis. We evaluated a frozen Apple Watch fusion system while making participant weighting, candidate provenance, and development history explicit. Methods: This post hoc exploratory participant-held-out internal validation used 46 adults, 249 nights, and 212,790 reference-valid 30-s epochs from BIDSleep. Four representations of instantaneous heart rate and accelerometry supplied an authoritative 47-candidate pool (36 final-stage and 11 carried-forward candidates). Within each frozen fold, a validation gate had excluded that fold’s outer-test labels; however, earlier staged development had inspected outer-fold outcomes and may have influenced the library. The primary estimand was participant-balanced four-class macro-[Formula: see text]. Confidence intervals used 50,000 participant bootstrap resamples; paired comparisons used 1,000,000 sign flips with Holm correction. Results: Participant-balanced macro-[Formula: see text] was 0.6813 (95% CI 0.6569–0.7036). The mean of five pooled outer-fold macro-[Formula: see text] values was 0.6964, accuracy was 70.76%, and weighted [Formula: see text] was 0.7067. Relative to the validation-selected single candidate from the same pool, sparse fusion increased participant macro-[Formula: see text] by 0.0255 (95% CI 0.0169–0.0346; 40/46 participants favored fusion; Holm [Formula: see text]). The gate activated 2–10 candidates per fold. Excluding all 11 carried-forward candidates yielded participant-balanced macro-[Formula: see text] 0.6804. A uniform average over all 47 candidates achieved a numerically similar participant-balanced macro-[Formula: see text] of 0.6833, but required all candidates and its paired difference from sparse fusion was uncertain. In a secondary common-support analysis restricted to the first 600 epochs of each night, raw fusion exceeded the SLAMSS reconstruction by 0.1066 in participant-level macro-[Formula: see text] (95% CI 0.0905–0.1222). Conclusions: Frozen sparse fusion outperformed a validation-selected single candidate and retained similar numerical performance after removing earlier-stage candidates. These exploratory internal results are competitive but do not establish cross-study superiority or external generalizability.