Non-Obtrusive Monitoring of Obstructive Sleep Apnea Syndrome Based On Ballistocardiography: A Preliminary Study

Biyong Zhang ^1,2 Zheng Peng ³ Chunjiao Dong ⁴ Jun Hu ¹ Xi Long ³ Tan Lyu ⁵ Peilin Lu ^6*

• ¹ Department of Industrial Design, Eindhoven University of Technology, Eindhoven, Netherlands
• ² Bobo Technology Ltd, Zhejiang, China
• ³ Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands
• ⁴ Hebei Medical University, Shijiazhuang, Hebei Province, China
• ⁵ Department of Cardiology, Sir Run Run Shaw Hospital, School of Medicine, Graduate School, Zhejiang University, Hangzhou, Jiangsu Province, China
• ⁶ Department of Neurology, Sir Run Run Shaw Hospital, School of Medicine, Graduate School, Zhejiang University, Hangzhou, Zhejiang Province, China

Obstructive sleep apnea syndrome (OSAS) degrades sleep quality and is associated with serious health conditions. Instead of the gold-standard polysomnography requiring complex equipment and expertise, a non-obtrusive device such as ballistocardiography (BCG) is more suitable for homebased continuous monitoring of OSAS, which has shown promising results in previous studies. However, often due to the limited storage and computing resource, also preferred by venders, the high computational cost in many existing BCG-based methods would practically limit the deployment for home monitoring. In this preliminary study, we propose an approach for OSAS monitoring using BCG signals. Applying fast change-point detection to first isolate apnea-suspected episodes would allow for processing only those suspected episodes for further feature extraction and OSAS severity classification. This can reduce both the data to be stored or transmitted and the computational load. Furthermore, our approach directly extracts features from BCG signals without employing a complex algorithm to derive respiratory and heart rate signals as often done in literature, further simplifying the algorithm pipeline. Apnea-hypopnea index (AHI) is then computed based on the detected apnea events (using a random forest classifier) from the identified apnea-suspected episodes. To deal with the expected underestimated due to missing true apnea events during changepoint detection, we apply boundary adjustment on AHI when classifying severity. Cross-validated on 32 subjects, the proposed approach achieved an accuracy of 71.9% for four-class severity classification and 87.5% for binary classification (AHI less than 15 or not). These findings highlight the potential of our proposed BCG-based approach as an effective and accessible alternative for continuous OSAS monitoring.