Wearable sleep staging performance declines with sleep apnea severity and is sensitive to training-set composition in very severe sleep apnea

We aimed to evaluate wearable sleep staging across sleep apnea severity, including very severe sleep apnea defined as an apnea-hypopnea index (AHI) [≥] 50 events/h, and to assess how training-set composition affects performance in this subgroup. We analyzed 552 overnight recordings: 318 from the Sleep Lab Dataset and 234 from the Hospital Dataset, of which 26.5% (N=62) had very severe sleep apnea. A deep learning model performed sleep staging from photoplethysmography-derived RR intervals and accelerometry recorded by a wrist-worn device. Baseline performance was assessed by 4-fold cross-validation using randomly partitioned folds from the combined datasets. We examined night-level associations with AHI severity. We also compared the baseline model with an ablation model trained on the same number of recordings but with all Sleep Lab Dataset recordings and lower-AHI Hospital Dataset recordings, evaluating both in the very severe subgroup. For 5-stage classification, Cohen's kappa was 0.586 in the Sleep Lab Dataset and 0.446 in the Hospital Dataset. Under 4-stage staging, the gap narrowed, with kappa values of 0.632 and 0.525, respectively. In the Hospital Dataset, kappa declined with AHI severity, with median kappa differing by about 0.2 between mild and very severe groups. In the very severe subgroup, kappa decreased from 0.365 (baseline) to 0.303 (ablation). Wearable sleep staging performance tended to decline across greater sleep apnea severity. Clinical utility may benefit from training data spanning the target severity spectrum and staging granularity matched to the intended use.