Sleep data from wearable device may help predict preterm birth
WashU interdisciplinary research uses machine learning to offer accurate prediction of preterm birth
Preterm birth complications are the primary cause of death among children under age 5, and nearly 75% could be prevented with interventions, according to the World Health Organization. While the causes leading to preterm birth are largely unknown, an interdisciplinary research team at Washington University in St. Louis has found that variability in sleep patterns in people experiencing pregnancy can effectively predict preterm birth.
Ben Warner, a doctoral student in the McKelvey School of Engineering, and Peinan Zhao, assistant professor of obstetrics & gynecology at WashU Medicine, used machine learning models to analyze sleep data from participants experiencing pregnancy. While disrupted sleep is known as a predictor of preterm birth, which is delivery before 37 weeks’ gestation, the reasons behind it have been unclear because the data was self-reported by patients.
Warner, who is advised by Chenyang Lu, the Fullgraf Professor in the Department of Computer Science & Engineering in McKelvey Engineering; Zhao; and Sarah England, the Alan A. and Edith L. Wolff Professor and vice chair for research of obstetrics and gynecology at WashU Medicine, studied data from 665 patients in the first two trimesters of their pregnancy with a recorded delivery date. The patients were in a 2014 cohort study conducted at Washington University in St. Louis and BJC HealthCare. About 14% of patients in the cohort experienced a preterm birth.
The patients wore a clinically validated wristwatch, called an actigraph, that measured body movements for roughly two-week periods. The data allows the team to extract daily patterns in the length of sleep, what time the patients went to sleep and woke up, their movement during sleep and several other variables. Patients also completed surveys about their sleep patterns. Warner and Zhao combined the two sources of data and plugged them into machine learning models to learn the impact of sleep patterns on preterm birth.
“We found that measures of sleep are decently predictive of preterm birth,” Warner said. “Variability in sleep patterns tends to be a stronger predictor of preterm birth than average sleep metrics, and getting consistent sleep is more important than getting good sleep on average.”
Lu’s lab has studied data from wearable devices for a wide variety of issues including contact tracing for COVID-19, predicting surgical outcomes and detecting depression and anxiety.
“Raw data from wearables can be very messy, but using a healthy combination of statistical methods, AI and clinical knowledge, researchers can extract important clinical insights,” said Lu, also director of the university’s AI for Health Institute. “Then AI scientists and clinicians work together to extract the insights from these very complex data from the real world and get meaningful insights from it.”
Zhao said their models were intentionally simple and looked for clinically significant association within the results.
“Our model shows that a machine learning model is better than a more statistical model,” said Zhao, who also is a member of the Center for Reproductive Health Sciences. “We can look at the result on importance of individual variables: How much does this variable contribute to the predictive model, and how much does it affect the final result? Based on that, a direction of a potential intervention is to promote a more consistent sleep schedule.”
Although many people experiencing pregnancy report disrupted sleep, that generally comes in the third trimester, England said. The team focused on finding signals before women were 20 weeks pregnant.
“There is no intervention because we can’t predict who’s going to have a preterm birth,” said England, director of the Center for Reproductive Health Sciences. “We’re hoping that this will be much more helpful in getting predictive power of women who are going to be at higher risk.”
Looking ahead, the team plans to validate the results in other populations at other academic medical centers.
“This research highlights the collaborations between engineering and obstetrics and gynecology,” England said. “Many people don't see natural connections between engineering and the field of reproduction, and this is another example of a perfect way for engineers to interact with researchers in our field.”
Warner BC, Zhao P, Herzog ED, Frolova AI, England SK, Lu C. Validation of sleep-based actigraphy machine learning models for prediction of preterm birth. npj Women’s Health. June 20, 2025. DOI: 10.1038/s44294-025-00082-y.
Funding for this research was provided by the March of Dimes Foundation, the Fullgraf Foundation, St. Louis Children's Hospital, Barnes-Jewish Hospital, the Department of Obstetrics and Gynecology at Washington University School of Medicine in St. Louis, Washington University AI for Health Institute, and the James McKelvey School of Engineering at Washington University in St. Louis.
The code used in this study is available at https://github.com/bcwarner/mod-actigraphy-clf.
