Detection of sleep apnea using only inertial measurement unit signals from apple watch: a pilot-study with machine learning approach

• Published in: Sleep & breathing Vol. 29; no. 1; p. 91
• Main Authors: Hayano, Junichiro, Adachi, Mine, Murakami, Yutaka, Sasaki, Fumihiko, Yuda, Emi
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.03.2025; Springer Nature B.V
• Subjects: Adult; Aged; Algorithms; Apnea; Dentistry; Female; Heart rate; Humans; Hygiene; Internal Medicine; Learning algorithms; Machine Learning; Male; Medicine; Medicine & Public Health; Middle Aged; Neurology; Otorhinolaryngology; Pediatrics; Pilot Projects; Pneumology/Respiratory System; Polysomnography - instrumentation; Regression analysis; Respiration; Signal Processing, Computer-Assisted - instrumentation; Sleep apnea; Sleep Apnea Syndromes - diagnosis; Sleep Breathing Physiology and Disorders • Original; Sleep Breathing Physiology and Disorders • Original Article; Sleep disorders; Seismocardiogram; Wearable sensor; Gyrocardiogram; Apnea-hypopnea index; Gyroscope; Home sleep apnea testing; Acceleration; Out-of-center sleep testing
• ISSN: 1520-9512; 1522-1709; 1522-1709
• DOI: 10.1007/s11325-025-03255-w

Purpose Despite increased awareness of sleep hygiene, over 80% of sleep apnea cases remain undiagnosed, underscoring the need for accessible screening methods. This study presents a method for detecting sleep apnea using data from the Apple Watch’s inertial measurement unit (IMU). Methods An algorithm was developed to extract seismocardiographic and respiratory signals from IMU data, analyzing features such as breathing and heart rate variability, respiratory dips, and body movements. In a cohort of 61 adults undergoing polysomnography, we analyzed 52,337 30-second epochs, with 12,373 (23.6%) identified as apnea/hypopnea episodes. Machine learning models using five classifiers (Logistic Regression, Random Forest, Gradient Boosting, k-Nearest Neighbors, and Multi-layer Perceptron) were trained on data from 41 subjects and validated on 20 subjects. Results The Random Forest classifier performed best in per-epoch respiratory event detection, achieving an AUC of 0.827 and an F1 score of 0.572 in the training group, and an AUC of 0.831 and an F1 score of 0.602 in the test group. The model’s per-subject predictions strongly correlated with the apnea-hypopnea index (AHI) from polysomnography ( r  = 0.93) and identified subjects with AHI ≥ 15 with 100% sensitivity and 90% specificity. Conclusion Utilizing the widespread availability of the Apple Watch and the low power requirements of the IMU, this approach has the potential to significantly improve sleep apnea screening accessibility.