Abstract Introduction Advancements in wearable devices have made it possible to easily and noninvasively acquire biological signals over extended periods. However, when using wearable devices to evaluate sleep in individuals with or without sleep disorders, it is necessary to examine how overall device performance varies with the type and severity of the disorder. Methods We developed a deep learning model that uses RR intervals and acceleration data measured by a wrist-worn device as input and examined its performance in sleep stage classification. Simultaneous measurements were conducted using polysomnography (PSG) and the wrist-worn device on healthy individuals and patients with sleep apnea (SA), and PSG assessments by specialists were used as the reference standard. The evaluation included two groups—healthy individuals (50 nights) and patients with SA (50 nights)—with accuracy and Cohen’s kappa calculated for 5-stage classification, and the mean absolute error (MAE) for estimating total sleep time (TST). Furthermore, the SA group was divided into three subgroups based on AHI: (I) AHI 30 (14 nights), (II) 30≤AHI 50 (21 nights), and (III) AHI≥50 (15 nights), and each subgroup was evaluated separately. Results For the healthy group, the accuracy was 0.70, the kappa was 0.58, and the MAE was 14 min. For the SA group, the accuracy was 0.59, the kappa was 0.44, and the MAE was 26 min. Across the AHI-based subgroups, (I) showed 0.62/0.49/24 min, (II) 0.57/0.42/27 min, and (III) 0.60/0.42/29 min for the accuracy, kappa, and MAE, respectively. Conclusion Although performance in the SA group was slightly reduced compared with the healthy group, the MAE remained around 26 min (~10% of TST). In the AHI-based subgroups, (II) and (III) showed a modest reduction in accuracy and kappa compared with (I), while no marked difference was found between (II) and (III). The MAE for all subgroups stayed within 24–29 min, indicating limited influence of AHI. These findings suggest that even for patients with SA, regardless of severity (including very severe cases with AHI≥50), this method has practical potential as a convenient approach for sleep evaluations. Support (if any)
Ogaki et al. (Fri,) studied this question.