Combining Wearables with Nearables: Using a Multi-Device Machine Learning Approach Improves Sleep Tracking at Home

Document Type

Conference Proceeding

Publication Date

5-19-2025

Publication Title

Sleep

Abstract

Introduction: Wearables have expanded access to sleep data, but proprietary algorithms are inaccurate and legacy actigraphy algorithms are outdated. Furthermore, legacy algorithms were trained on nighttime sleep, and only identify 50.3% of daytime sleep. This presents a unique challenge for night shift workers, who are prone to disordered sleep and need access to improved sleep measurement tools. We recently found machine learning (ML) algorithms using raw accelerometer and heart rate data from an Apple Watch could improve nighttime sleep tracking achieving up to 90% accuracy. However, wearables alone still do not capture environmental inputs needed to accurately clas sify sleep versus wake (e.g. sleep onset latency). We conducted a proof-of-concept study assessing the feasibility of a ML approach that combines inputs from nearables and a wearable to improve daytime sleep tracking at home. Methods: Researchers installed a curated sleep tracking system in participants’ bedrooms to continuously monitor activity and the environment for 30 days. This included a presence sensor to detect presence in bed, a luxmeter to measure changes in ambi ent light, and a wireless light switch with smart light bulbs to track when lights were turned on and off. We also collected raw heart rate and accelerometer data from Apple Watches and raw accelerometer data from iPhones. Participants completed daily sleep diaries and a user experience questionnaire. Results: Preliminary results show our multi-device ML approach increases detection of daytime sleep by 43.4% in night shift workers. In nighttime sleepers, our ML approach achieves 93.7% sensitivity for sleep identification, while maintaining 97.2% specificity in wake classification. Participants report strong acceptance of the multi-device approach, with low per ceived intrusiveness (1.0 of 10) and high willingness to continue use (8.8 of 10). Conclusion: These findings support the feasibility of a multi-de vice ML approach for more accurate sleep tracking outside of the lab. We plan to expand this research to a larger sample of night shift workers to improve the precision of daytime sleep tracking, while assessing the sleep environment. Because our sleep tracking system contains smart technology, we ultimately aim to help inform personalized interventions to the bedroom environment to improve sleep outcomes.

Volume

48

First Page

A192

Last Page

A193

Find It @ Sladen

Share

COinS