For all their potential to drive changes in health, wearables have struggled to gain a foothold in medicine. The dramatic changes during pregnancy are a fertile ground to test their potential, though — and new research shows how applying machine learning methods to streams of data from wearable devices could be used to understand the mystery of premature birth.
Machine learning researchers at Stanford University used a deep learning model to analyze wearable activity and sleep data from pregnant participants. No surprise: Their sleep typically got worse and their activity slowed down over the course of pregnancy. But some participants had data profiles that didn’t match their pregnancy stage — and it was those pregnancies, the researchers found, that were more likely to result in a preterm birth.
Preterm birth is the leading cause of death in children under five around the world, and in the United States, about 11% of all live births occur before 37 weeks of gestation — a number that’s been steadily increasing over the last decade. Black women are particularly at risk, being about 1.5 times as likely as white women to deliver prematurely.
The new research, published Thursday in npj Digital Medicine, stands out among studies of wearable data because of the size and diversity of its participant group. More than 1,000 women from the St. Louis area, more than half of whom were Black, were followed through their pregnancies by researchers from Washington University in St. Louis.
“They really took on this monstrous effort to do this throughout the entire course of pregnancy,” said Nima Aghaeepour, a machine learning researcher at Stanford and senior author on the paper. The team tried to collect data as soon as possible after a woman became pregnant, and asked them to wear a motion- and light-sensing watch for at least a week during each trimester. Some wore it the whole nine months of their pregnancy.
Aghaeepour and his colleagues fed that raw, continuous data into a deep learning pipeline and analyzed it alongside the patients’ medical records — including how far along they were in their pregnancy. “It turns out that we can look at the wearable device and tell how pregnant somebody is, give or take a few weeks,” said Aghaeepour, based on their typical patterns of physical activity and sleep.
The participants’ medical records followed them all the way to birth — allowing the researchers to see who gave birth early and whose pregnancies went to term.
“The place where things get exciting,” said Aghaeepour, “is that there are these women who are not very pregnant — their gestational age is low — but they look very pregnant to the deep learning algorithm.” Their patterns of activity and sleep were more disturbed than women at similar stages of pregnancy. It’s those women, the researchers found, who were about 44% more likely to give birth prematurely.
That result doesn’t mean that lack of activity or sleep is causing more preterm births. “We are not saying that we can estimate when your baby is going to be born based upon your Fitbit data,” said Erik Herzog, a circadian biologist at Washington University in St. Louis and co-author on the paper. But it does suggest a hypothesis about the role of activity and sleep in premature birth that can be tested in future research.
“Next time, when we have a bunch of women wearing these things and generating data, let’s see if we can generate alerts that say, this person looks not like they’re approaching week 40, but like they’re approaching delivery,” said Benjamin Smarr, a professor at the University of California San Diego who has studied the use of temperature data from the wearable Oura ring to detect pregnancy. “Is it the case that we’re guessing that accurately? Is it the case that if we intervene early we end up catching some of them?”
That kind of prospective analysis could also help find explanations for racial disparities in preterm birth. Researchers have struggled to identify what environmental and societal factors drive poorer outcomes for Black women. “Where are these complications coming from? Because it’s not just access to education, it’s not just socioeconomic status,” said Smarr.
Pregnancy is uniquely suited for this kind of research, said Jessica Walter, a reproductive endocrinologist at Northwestern who researches the applications of wearables in women’s health. “You have extremely motivated individuals wanting to understand their health, to take care of themselves, during a really dynamic period of changing physiology during a pregnancy,” she said.
That’s led other research groups to try to uncover signals about pregnancy with wearable devices, including an ambitious study led by the nonprofit 4YouandMe, co-founded by former Apple Health researcher Stephen Friend. The Better Understanding the Metamorphosis of Pregnancy study — BUMP, for short — has set out to test the feasibility of collecting hundreds of variables from about 1,000 pregnant participants using commercial devices, and to describe the variability in their values.
These research efforts are confronting a crucial challenge of working with continuous data from wearable devices. Their output is so variable — both between individuals, and across an individual’s day-to-day fluctuations — that it’s difficult to pull out actionable information from simple measures like average activity or daily sleep length. “[Machine learning] is one tool that we can use to help us manage these extremely dense, large volume datasets,” said Walter, uncovering less obvious patterns that carry meaningful signals about health.
“It really goes at the data in an unbiased way,” said Herzog. “This machine learning approach said, we’re just going to teach the model that these data are associated with these gestational ages, and it learned features in the data that allowed it to better estimate gestational age.”
It’s not enough simply to trust that the machine gets things right, of course. “The deep learning model remains kind of a black box, which is always a little bit of a challenge in health,” said Smarr. And if wearables can indeed be used to catch patients at risk of preterm birth, those methods still need to be tested carefully in the real world. “Do you stress the person out, do you cost the hospital resources? There’s ultimately these tradeoffs.”