Performance of a Smart Ring for Sleep Staging and Sleep Apnea Screening: A Validation Study with Inter-Generation Comparison (Preprint)

<sec> <title>BACKGROUND</title> <p>Consumer wearables are increasingly used to assess sleep, but their accuracy across device generations remains unclear. Using a standardized framework to validate a device facilitates comprehensive assessment of its performance and provides a reliable reference for its subsequent application and research.</p> </sec> <sec> <title>OBJECTIVE</title> <p>We aimed to evaluate the accuracy of a smart ring for sleep staging and sleep apnea screening, as well as inter-generation comparison, using a standardized validation framework for wearable sleep-monitoring devices, with PSG as the reference standard.</p> </sec> <sec> <title>METHODS</title> <p>Adults from the South China Sleep Health Cohort underwent concurrent overnight PSG and RingConn smart ring monitoring. Sleep metrics and 30-second sleep-stage classification included 306 participants and 305,615 epochs; apnea–hypopnea index (AHI) analyses included 252 participants. Data were collected from June 2024 to June 2026. Agreement was assessed using intraclass correlation coefficients (ICCs) and Bland–Altman analyses, sleep-stage classification using sensitivity, specificity, and kappa statistics, and OSA screening at AHI thresholds of 5, 15, and 30 events/hour.</p> </sec> <sec> <title>RESULTS</title> <p>Among 306 participants (median age=37 years (interquartile range: 16); 86 (28%) female), mean ring-minus-PSG differences were 14.01 min for total sleep time, 2.63% for sleep efficiency, 5.57 min for sleep onset latency, 11.07 min for rapid eye movement (REM) sleep duration, 24.81 min for deep sleep duration, -19.58 min for wake after sleep onset, and -21.78 min for light sleep. The proportional bias was evident across all metrics. Sensitivity was 93% for sleep detection, 54% for wake, 58% for light sleep, and 40%, for both deep and REM sleep, with similar performance between device generations. For AHI, the ICC was 0.85 (95% confidence interval: 0.81-0.88) and the mean difference was -9.14 (-10.50 to -7.78) events/hour, becoming more negative at higher PSG-derived AHI. Sensitivities for classifying normal, mild, moderate, and severe OSA were 93.8%, 31.9%, 36.6%, and 57.1%, respectively. At thresholds of 5, 15, and 30 events/hour, the area under curve (AUC) was 0.85, 0.86, and 0.78, respectively.</p> </sec> <sec> <title>CONCLUSIONS</title> <p>The smart ring demonstrated high sensitivity for sleep detection and potential for OSA screening, but limited accuracy for wake detection, four-class sleep staging and OSA severity grading compared with PSG.</p> </sec> <sec> <title>CLINICALTRIAL</title> <p>This study was approved by the School of Public Health, Sun Yat-sen University (approval No. [2024]108) and registered with the China Clinical Trial Registry (No. ChiCTR2600121781).</p> </sec>