Belun Ring (Belun Sleep System BLS-100): Deep learning-facilitated wearable enables obstructive sleep apnea detection, apnea severity categorization, and sleep stage classification in patients suspected of obstructive sleep apnea

• Published in: Sleep health Vol. 9; no. 4; pp. 430 - 440
• Main Authors: Strumpf, Zachary, Gu, Wenbo, Tsai, Chih-Wei, Chen, Pai-Lien, Yeh, Eric, Leung, Lydia, Cheung, Cynthia, Wu, I-Chen, Strohl, Kingman P., Tsai, Tiffany, Folz, Rodney J., Chiang, Ambrose A.
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.08.2023
• Subjects: Apnea-hypopnea index; Artificial intelligence; Digital health; Home sleep apnea testing; Obstructive sleep apnea; Peripheral arterial tonometry; Photoplethysmography; Sleep technology; Validation; Validation; Sleep technology; Apnea-hypopnea index; Photoplethysmography; Digital health; Home sleep apnea testing; Obstructive sleep apnea; Artificial intelligence; Peripheral arterial tonometry
• ISSN: 2352-7218; 2352-7226; 2352-7226
• DOI: 10.1016/j.sleh.2023.05.001

Our objective was to evaluate the performance of Belun Ring with second-generation deep learning algorithms in obstructive sleep apnea (OSA) detection, OSA severity categorization, and sleep stage classification. Belun Ring with second-generation deep learning algorithms In-lab polysomnography (PSG) Eighty-four subjects (M: F = 1:1) referred for an overnight sleep study were eligible. Of these, 26% had PSG-AHI<5; 24% had PSG-AHI 5–15; 23% had PSG-AHI 15–30; 27% had PSG-AHI ≥ 30. Rigorous performance evaluation by comparing Belun Ring to concurrent in-lab PSG using the 4% rule. Pearson’s correlation coefficient, Student’s paired t-test, diagnostic accuracy, sensitivity, specificity, positive predictive value, negative predictive value, positive likelihood ratio, negative likelihood ratio, Cohen’s kappa coefficient (kappa), Bland-Altman plots with bias and limits of agreement, receiver operating characteristics curves with area under the curve, and confusion matrix. The accuracy, sensitivity, specificity, and kappa in categorizing AHI ≥ 5 were 0.85, 0.92, 0.64, and 0.58, respectively. The accuracy, sensitivity, specificity, and Kappa in categorizing AHI ≥ 15 were 0.89, 0.91, 0.88, and 0.79, respectively. The accuracy, sensitivity, specificity, and Kappa in categorizing AHI ≥ 30 were 0.91, 0.83, 0.93, and 0.76, respectively. BSP2 also achieved an accuracy of 0.88 in detecting wake, 0.82 in detecting NREM, and 0.90 in detecting REM sleep. Belun Ring with second-generation algorithms detected OSA with good accuracy and demonstrated a moderate-to-substantial agreement in categorizing OSA severity and classifying sleep stages.