Human health is directly tied to sleep stages, and physical and mental well-being is dependent on cycling through them, she says. Disruptions can increase risks for chronic diseases, poor cognition, and mood issues.
The sleep-staging methodology is intended to “extract more robust performance” from a ubiquitous, non-specialized consumer wearable device to make it possible to monitor people with sleep disorders at home, says Dutta. Capabilities of the framework were described in a research article where 47 healthy adults recorded their sleep for up to seven consecutive nights using an Apple Watch Series 6 and a Dreem 2 Headband (IEEE Transactions on Biomedical Engineering, DOI: 10.1109/TBME.2025.3612158).
The underlying architecture is based on a long short-term memory (LSTM) deep learning model commonly used for time series analysis whereby sequential data is processed by learning to remember or forget information over long periods. Sleep is staged using instantaneous heartrate and accelerometry.
Since the model was previously validated in two larger study populations (PLOS One, DOI: 10.1371/journal.pone.0285703), it had a solid starting point, Dutta says. The latest study presents ways to customize the sleep staging approach by optimizing the input data collected using wrist wearables such as the Apple Watch, including mean heartrate and heartrate variability in a 30-second period.
Dutta’s primary research interest is the connection between sleep disruptions and Alzheimer’s disease, which she is tackling from multiple angles. This involves working with PET and MR imaging datasets and techniques, the collection of blood biomarkers of Alzheimer’s disease and, more broadly, how early changes in sleep patterns relate to people’s path of decline toward developing dementia.
Many different trajectories are possible now that the basic app has been created, says Dutta. “It doesn’t have to be dementia; it could be something else where sleep is relevant,” such as clinical trials looking at mood disorders or the sleep effects of various medical procedures and therapies. Once validated for a particular disease or cohort, BIDSleep might find utility as a dedicated research product for different groups of individuals to relate how they are sleeping with a clinical endpoint associated with their condition.
The consumer market is now flooded with portable sleep-tracking devices, some looking at brain activity through electroencephalography (EEG) patches or headbands, but all far more accessible and convenient than technically complex, lab-based polysomnography involving a high number of sensors simultaneously recording up to 20 different physiological parameters, Dutta says. While obviously not as accurate as the gold standard test, “for the average person who is interested in tracking sleep or having some sleep disruptions ... it’s important to have options that enable people to individually look at their own sleep health.”
Unless they’re “super-curious,” she adds, they likely wouldn’t want to invest in a dedicated device. This is what gives all-purpose, wrist-based solutions the edge for purposes of assessing their sleep health. “There is a strong motivation to use these devices just because of how popular they are, and this could make a lot of rich sleep information more accessible to a large number of individuals.”
Dutta and her colleagues picked the Apple Watch as their target device because of its sheer popularity relative to the many alternative Android devices. The Apple Watch now has a built-in sleep assessment tool, a feature unavailable at the time of their study, although it provides information that is understandably a “bit coarser” than their sleep-staging model paired with the custom app since the company’s goal is to optimize the overall device—multiple functionalities, user interface, and battery life—and not just the collection of sleep-relevant data.
Researchers trying to study sleep can use the hardware capabilities of the Apple Watch, just “in a slightly different way where they ... take more frequent looks at sleep-relevant datasets that the watch is already collecting and then processing all of that using an advanced AI model that can interpret this data and figure out the proportions of light, deep, and REM sleep as [someone is] sleeping.”
The model comprises multi-layered neural networks that are associated with “lots of unknown parameters” that it learns during its pre-validation training phase, she explains. It is then tested on datasets where the parameters are frozen to assess how accurately it predicts whatever the endpoint is for the problem at hand.
To maintain consistency in running the recent Apple Watch study led by Dutta, participants were provided with the devices to ensure they were all using the same version. They were also given an iPhone, which acts as the primary interface for data management.
On average, the AI-based model built in Dutta’s lab accurately identified the correct sleep stage 71% of the time, outperforming other well-known approaches used by the sleep research community, she reports. It was also found to be more accurate at identifying deep sleep, a pronounced decline of which is associated with aging more so that the amount of total sleep measured by smartwatches.
The research team did a two-part release of the results, Dutta explains, starting with an “ablation study” where they systematically removed components of the overall model to see how much each of them contributed to its overall performance. The goal here was to identify the optimal network architecture.
They then made an extensive set of comparisons with five alternative sleep-tracking methods. These included three well-known models in the time series data analysis field as well as two more basic models focused on instantaneous heartrate that they pushed to their highest potential to ensure a fair benchmarking exercise against their own optimized LSTM model.
The model’s capabilities include the measurement of sleep efficiency and sleep onset latency (how long it takes to fall asleep), both of which tend to be less reliably captured by smartphones and most especially manual self-reporting in sleep questionnaires, she adds. As a practical matter, sleep logs are used in large-cohort studies and may produce a statistical effect, “but at the individual level there might be quite a bit of variability.”
In trying to connect sleep to dementia, researchers must consider many other traits influencing disease risk, Dutta points out. “In one of our cohorts, for instance, we are looking at individuals who are at elevated genetic risk of dementia and in another cohort, we are just looking at subjects who have a certain number of known risk factors for Alzheimer’s disease.”
The broader goal in her lab is to establish how disruptive sleep patterns relate to cognitive impairment, and specifically how this happens via studies in early-stage individuals, she continues. Despite the many knowledge gaps in the Alzheimer’s field, “there is general consensus that all these therapeutics being developed need to be administered at ... the preclinical phase where you don’t necessarily have many [or any] memory complaints ... but you may have some of these signature pathologies of Alzheimer’s disease, primarily amyloid and tau.”
These abnormal proteins often accumulate in the brain early in the disease process when people also tend to have disrupted sleep patterns, says Dutta, in explaining the rationale for studying the phenomenon in larger cohorts using smartwatches. Studies she is currently conducting involve using portable EEG devices to look at brain activity in unison with Apple Watches to measure instantaneous heart rate and accelerometry.
The work will involve extending the group’s latest benchmarking exercise to the Apple Watch’s native sleep-staging capabilities, she adds. Follow-up studies with early-stage dementia patients using the wearables will investigate how different pathological measures relate to sleep-based measures and manifest in cognitive changes downstream.
“We have a broader interest in using multiple wearables to gather rich information about sleep, which would give us a multimodal perspective on sleep tracking,” she says. “With that goal in mind, we’re interested in seeing how sleep patterns in very early, pre-dementia phases can be telling of future signs of cognitive impairment, and to do that at a larger scale.”
The code used to produce sleep staging results in the latest study has been posted online, and the BIDSleep app enabling the transmission of collected sleep data to any email address is available in the Apple Store, reports Dutta. The dataset is also in the process of being publicly released, enabling other researchers to similarly train a model to get sleep staging results.
The AI model is a prototype and utilizing it requires some domain knowledge, she says. But, pending the outcome of ongoing dementia studies, the team is interested in adding a user interface to make the tool more broadly usable so that even people with no AI expertise could run the model on their own dataset to find out how they’ve been sleeping.
Dutta says she has been fascinated with sleep science over the last five or six years and is increasingly convinced of its importance to the health of people of every age. The “future is bright” in terms of the capabilities of many wearable devices, coupled with AI, to empower individuals to track their sleep and overall health outside of the clinic setting, she adds.