1206 Automated Sleep Apnea Assessment Based On Machine Learning And Wearable Technology

• Published in: Sleep (New York, N.Y.) Vol. 43; no. Supplement_1; p. A461
• Main Authors: Chiu, H, Lee, P, Ku, B, Liu, Y
• Format: Journal Article
• Language: English
• Published: US Oxford University Press 27.05.2020
• Subjects: Automation; Machine learning; Sleep apnea; Wearable computers
• ISSN: 0161-8105; 1550-9109
• DOI: 10.1093/sleep/zsaa056.1200

Abstract Introduction Obstructive sleep apnea (OSA) is a condition characterized by repeated episodes of partial or complete obstruction of the respiratory passages during the sleep. Traditional polysomnography (PSG) for OSA estimation is bulky and time-consuming for daily use. Therefore, this study aims to develop a novel photoplethysmography (PPG) and accelerometer based smart watch for OSA detection, in which a high-performance and low-complexity automated OSA detection was embedded for long-term in-home measurement. Methods The developed watch measured PPG signals from wrist radial artery and body motion from accelerometer as well. 121 patients (92 males, 29 females) were recruited from the normal community and Center of Sleep, National Taiwan University Hospital, Taiwan in this study. All OSA scoring were analyzed by three registered PSG technologists. The AHI of the cohort was 10.1±18.3 (0 to 82.7). An automated OSA detection algorithm was designed based on machine-learning (ML) technique, in which was iteratively updated according to each 30-second epoch of the collected data. Subsequently, obstructive and hypopnea events were detected according to the OSA detection algorithm. Results To better valid the effectiveness, this study focused on the estimation performance of the subjects with AHI>15. Based on hold-out validation, the average sensitivity and precision in the AHI>15 cohort were 77.2% and 58.6%, respectively, with a Cohen’s kappa of 0.46. The interclass correlation between the watch and technologists was 0.81 (95%CI: 0.61-0.91). The result showed that the proposed automated OSA detection could achieve consistent result with technologist during standard sleep testing. Conclusion This study developed a wrist-based watch based on ML technique to assess OSA severity. We compared the performance with clinical technologists for the OSA detection. Further, the sensitivity and precision were generally acceptable while subjects were with AHI>15. The proposed wrist-based watch could provide reliable performance for OSA estimation, and may be of a light for future in-home sleep studies. Support This study was supported by Mediatek Inc.