Related Experiment Video
Updated: Sep 22, 2025
Assessing the Accuracy of Fitness Smartwatch Data for Cardiovascular and Physical Activity Monitoring: A Validation Study in Digital Health
Published on: February 21, 2025
Obstructive Sleep Apnoea Syndrome Screening Through Wrist-Worn Smartbands: A Machine-Learning Approach
Davide Benedetti1, Umberto Olcese2, Simone Bruno1
1Department of Translational Research and of New Surgical and Medical Technologies, University of Pisa, Pisa, Italy.
Purpose:
A large portion of the adult population is thought to suffer from obstructive sleep apnoea syndrome (OSAS), a sleep-related breathing disorder associated with increased morbidity and mortality. International guidelines include the polysomnography and the cardiorespiratory monitoring (CRM) as diagnostic tools for OSAS, but they are unfit for a large-scale screening, given their invasiveness, high cost and lengthy process of scoring. Current screening methods are based on self-reported questionnaires that suffer from lack of objectivity. On the contrary, commercial smartbands are wearable devices capable of collecting accelerometric and photoplethysmographic data in a user-friendly and objective way. We questioned whether machine-learning (ML) classifiers trained on data collected through these wearable devices would help predict OSAS severity.
Patients And Methods:
Each of the patients (n = 78, mean age ± SD: 57.2 ± 12.9 years; 30 females) underwent CRM and concurrently wore a commercial wrist smartband. CRM's traces were scored, and OSAS severity was reported as apnoea hypopnoea index (AHI). We trained three pairs of classifiers to make the following prediction: AHI <5 vs AHI ≥5, AHI <15 vs AHI ≥15, and AHI <30 vs AHI ≥30.
Results:
According to the Matthews correlation coefficient (MCC), the proposed algorithms reached an overall good correlation with the ground truth (CRM) for AHI <5 vs AHI ≥5 (MCC: 0.4) and AHI <30 vs AHI ≥30 (MCC: 0.3) classifications. AHI <5 vs AHI ≥5 and AHI <30 vs AHI ≥30 classifiers' sensitivity, specificity, positive predictive values (PPV), negative predictive values (NPV) and diagnostic odds ratio (DOR) are comparable with the STOP-Bang questionnaire, an established OSAS screening tool.
Conclusion:
Machine learning algorithms showed an overall good performance. Unlike questionnaires, these are based on objectively collected data. Furthermore, these commercial devices are widely distributed in the general population. The aforementioned advantages of machine-learning algorithms applied to smartbands' data over questionnaires lead to the conclusion that they could serve a population-scale screening for OSAS.
