Discovering wearable sensor-derived temporal digital phenotypes and their association with hypertensive disorders in pregnancy
This study demonstrates that machine learning-derived digital phenotypes from wearable sensors, specifically capturing transitions toward lower physical activity, poorer sleep, and reduced autonomic variability, can effectively predict subsequent hypertensive disorders of pregnancy, offering a promising complement to routine prenatal care.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Pregnancy is a time of profound physiological change, yet the standard medical check-up offers only a fleeting glimpse into that transformation. A doctor measures blood pressure and asks how a patient feels, but these snapshots happen weeks apart, often missing the subtle, continuous shifts that occur between visits. For conditions like hypertensive disorders of pregnancy, which involve dangerous spikes in blood pressure and can lead to severe complications for both mother and baby, this gap in observation is a critical problem. Scientists have long suspected that the body gives early warning signs through changes in daily habits and internal rhythms long before a crisis occurs. The emerging field of digital phenotyping seeks to capture these continuous signals, using the data streams from everyday wearable devices to build a living picture of health. Instead of relying on a single moment in time, this approach looks for patterns in how a person moves, sleeps, and how their heart rate varies over days and weeks, hoping to spot the early drift toward illness.
In a new study, researchers at Johns Hopkins University explored whether these continuous streams of data could reveal specific, interpretable patterns of health that predict hypertensive disorders. They turned to a large group of pregnant participants in the NIH All of Us Research Program, who had worn Fitbit devices and whose medical records were available for review. The team focused on three key areas: physical activity, sleep, and pulse rate variability, a measure of how much the time between heartbeats changes, which reflects the nervous system's ability to adapt. Rather than looking at raw numbers, the researchers used a sophisticated computer model to group thousands of two-week periods of data into distinct clusters. They treated each two-week window as a single unit of time, analyzing the days leading up to the end of each trimester and the postpartum period to see if the body's rhythm had shifted.
The model successfully identified distinct "digital phenotypes," or behavioral and physiological profiles, for each of the three areas. For physical activity, the data sorted into three clear groups: one characterized by high movement and low sedentary time, a middle group, and a third group defined by low activity and high sedentary time. Sleep patterns separated into two main types: one group that slept longer with more restorative deep and rapid eye movement sleep, and another that slept less with more time spent awake or restless. Similarly, pulse rate variability grouped participants into those with higher variability, indicating a more flexible nervous system, and those with lower variability. These clusters were not random; they represented real, measurable differences in how the participants lived and how their bodies functioned.
The most significant finding emerged when the researchers tracked how participants moved between these groups over time. They discovered that a person's risk of developing a hypertensive disorder increased if their physiological profile worsened as the pregnancy progressed. Specifically, if a woman transitioned from a highly active state to a less active one, or from restorative sleep to fragmented sleep, her odds of developing high blood pressure complications rose. These transitions were most telling when they occurred between the second and third trimesters and as the pregnancy moved into the postpartum period. The data suggested that the body's loss of stability over time was a stronger warning sign than any single measurement taken at a clinic visit. Furthermore, the risk was highest for those who showed signs of decline across all three areas simultaneously, suggesting that these disorders arise from a broad, multi-system breakdown rather than a single isolated issue.
To ensure these findings were not unique to the specific devices or people in the first group, the team tested their activity model on a separate group of pregnant women from Washington University. Even though this second group wore different devices and followed a different protocol, the computer model still identified the same three activity clusters with similar clarity. The patterns held up, and the model's ability to distinguish between different activity levels remained strong. This external validation provided confidence that these digital signatures of health are robust and could potentially be recognized across different populations and sensing technologies.
The study also looked at which days within the two-week windows mattered most. For physical activity and sleep, the computer model paid the most attention to the final few days of each window, suggesting that short-term changes in behavior often precede a clinical diagnosis. In contrast, the model looked at the entire two-week period for pulse rate variability, indicating that changes in the nervous system are a slower, more cumulative process. These insights help explain how the body might be signaling distress: a sudden drop in movement or a few nights of poor sleep might be the immediate precursor to a hypertensive event, while the underlying autonomic changes build up more gradually.
While the study offers a promising new way to view pregnancy health, the authors are careful to note its limitations. The participants were a specific group who agreed to wear devices and share data, which may not represent all pregnant people. The study was also retrospective, meaning the researchers looked back at data that had already been collected, rather than testing the system in real-time to prevent complications. Additionally, the pulse rate data came from wrist sensors, which are less precise than medical-grade heart monitors. Despite these constraints, the work demonstrates that continuous monitoring can uncover hidden patterns of risk. It suggests that the future of prenatal care might involve not just periodic check-ups, but a continuous, passive stream of information that flags when a patient's daily rhythms begin to drift away from the healthy norm, offering a chance for earlier intervention.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.