AI-powered app turns consumer wearables into sleep study devices

An app that turns consumer Apple Watches into tools for highly sophisticated sleep stage monitoring was developed by team of researchers led by professor Joyita Dutta at the University of Massachusetts Amherst. The researchers say the app and corresponding AI code are convenient and effective alternatives to existing costly and complex sleep study equipment and protocols.

"Our goal was to get as rugged as possible with a non-specialized consumer wearable device, which is the Apple Watch," says Dutta, professor of biomedical engineering in the Daniel J. Riccio Jr. College of Engineering and senior author of the research, published in IEEE Transactions on Biomedical Engineering. She envisions researchers can use this app to monitor people with sleep disorders at home, without costly lab-based sleep studies.

Dutta designed this app specifically for her research into the connection between sleep disruptions and the development of Alzheimer's disease.

Currently, the gold standard for sleep studies is through lab-based assessments, which are complex, expensive and require manual data analysis by a specialist. Even at-home assessments can be complicated, asking participants to sleep with electrodes on their heads.

Because of the cost, complexity and discomfort, the duration of most sleep studies is only one night, meaning researchers don't have the benefit of analyzing data from multiple sessions over time. Dutta also notes that, for her ongoing Alzheimer's research, existing monitoring technology cannot capture sleep data from naps, which are largely unplanned. In contrast, the broad availability and round-the-clock wearability of smartwatches make them particularly well-suited for studying all forms of sleep.

With this in mind, Dutta and her team created software to turn the widely available Apple Watch into a robust sleep-staging technology. The app, called BIDSleep, collects data on instantaneous heart rate, since this measure varies depending on the sleep stage. Heart rate is slower during deep sleep and higher during more active periods, like REM sleep.

These data feed into the researchers' new AI model, which is available to other researchers.

On average, their model accurately identified the correct sleep stage 71% of the time, outperforming other well-known approaches used by the sleep research community. Dutta also notes that their model is even more accurate at identifying deep sleep, which is important because aging is associated with more pronounced decline in deep sleep than total sleep.

"Overall accuracy matters, but sometimes we also need to look at the clinical metrics like sleep efficiency and sleep onset latency, total sleep time," adds Tzu-An Song, a postdoctoral research fellow in Dutta's lab and first author on the paper. Accuracy along these measures provides further insights into the app's effectiveness at predicting clinically important sleep parameters.

Our method works better for basically all of these metrics."

Tzu-An Song, first author on the paper

The AI model, using data collected by BIDSleep, produced results closest to the gold standard of EEG-based sleep staging compared to other modeling approaches.

Dutta notes that they did not compare their technology to the Apple Watch's native sleep-staging capabilities because that feature was not available at the time of their study. They plan to conduct a full head-to-head comparison in the future. She is optimistic that their app will be more accurate because it provides richer data, collecting heart rate information at a denser rate than the native features built into Apple Health.

"Ultimately, we'd love for researchers and clinicians to use this app, which is why we created it in a style where you can easily port the data and get multi-night information out of it," says Dutta.

This research is supported by the Massachusetts AI and Technology Center for Connected Care in Aging & Alzheimer's Disease and the National Institutes of Health.