Summary: A medical wristband and a machine learning model detected disrupted sleep patterns roughly 24 hours before participants began shedding the influenza virus. Although the original study focused on influenza, the researchers suggest the same approach could help flag the onset of other infections, potentially including COVID-19. This wristband-based monitoring is not a diagnostic tool, but it may provide an early warning signal that prompts precautionary self-isolation.
Source: University of Michigan
Overview
A preprint study using data from a medical wristband indicates that changes in sleep detected by wearable sensors can predict the early stages of influenza infection—about one day before participants became contagious. The research, posted on arXiv and submitted to IEEE Transactions on Biomedical Engineering, applied a machine learning algorithm to multisensor wristband data to identify shifts in sleep and wake patterns caused by infection.
Study design and key findings
Researchers trained an unsupervised transfer learning algorithm on data from a medical wristband worn by 25 volunteers who were deliberately exposed to an H3N2 influenza strain during a human viral challenge study. The model infers sleep and wake states from multisensor signals and adapts when sleep behavior shifts due to illness. Of the eight participants who developed influenza and had usable sensor recordings, seven showed detectable sleep disruptions about 24 hours before viral shedding began.
The model achieved strong predictive performance in this controlled setting. Features derived from the detected sleep/wake periods were highly predictive of both infection status and the timing of infection onset, reporting area-under-the-curve (AUC) values indicating good discrimination. These results suggest that sleep-pattern disruption, as measured by wrist-worn devices, can serve as an early marker of respiratory infection.
Machine learning and wearable sleep monitoring
The algorithm combines a multivariate hidden Markov model with Fisher’s linear discriminant analysis and a domain-adaptation scheme that incrementally learns changing sleep dynamics. Importantly, it operates without prior labels for true sleep or wake states, making it suitable for real-world ambulatory monitoring where ground truth is often unavailable. By leveraging integrated multisensor processing and adaptive training, the method maintains robust sleep/wake detection even when illness alters normal sleep behavior.
Implications for public health and wearable technology
While the technique cannot diagnose a specific disease like COVID-19, it could provide an early-warning system for individuals and organizations. If similar predictive signals can be obtained from consumer smartwatches and fitness trackers, the approach could notify users—especially critical workers—when to consider precautionary self-isolation before they become contagious. During outbreaks where asymptomatic or atypical presentations are common, an early signal based on physiological changes such as sleep disruption may be particularly valuable for limiting spread.
Researchers emphasize that more data from diverse populations and devices are needed to refine predictive models. Expanding training datasets to include community-based smartwatch data could improve generalizability and help distinguish infection-related sleep changes from other causes of sleep disruption.
Study context and authorship
The lead corresponding author is Alfred Hero, John H. Holland Distinguished University Professor of Electrical Engineering and Computer Science and a professor of biomedical engineering and statistics. The preprint titled “An unsupervised transfer learning algorithm for sleep monitoring” lists Xichen She, Yaya Zhai, Ricardo Henao, Christopher W. Woods, Geoffrey S. Ginsburg, Peter X.K. Song, and Alfred O. Hero among the authors. The research highlights how adaptive multisensor algorithms can improve automated sleep assessment and support early detection of infection-related physiological changes.
About this neurotech research article
Source:
University of Michigan
Media contact:
Kate McAlpine – University of Michigan
Image:
Public domain image
Original research: Closed access. Preprint available as arXiv:1904.03720 titled “An unsupervised transfer learning algorithm for sleep monitoring”.
If validated across broader populations and consumer devices, sleep-based wearable monitoring could become a practical early-warning tool to supplement testing and clinical assessment during infectious disease outbreaks.