Emily Pickett, a doula in Louisville, Kentucky, is used to hearing hard truths from expecting mothers. Her job is to guide women through pregnancy, acting as confidante and supporter; understanding their deepest stressors—an abusive partner, a struggle with drugs—is important to ensuring healthy pregnancies. In the largely poor and black neighborhoods of Louisville’s West End, where Pickett leads a nonprofit network of doulas, it’s also part of preparing women for a specific threat: giving birth too soon.

Premature birth and its complications are the leading cause of death in infants. In the US, roughly one in 10 babies is born prematurely, or before 37 weeks, far higher than most other developed countries, and the rate is ticking upward. That national rate masks big disparities: Black women are more than 50 percent more likely to deliver prematurely than whites. But it’s impossible to say for sure if a woman will deliver early, limiting the ways that doctors can intervene.

That’s sparked new efforts to identify and diagnose risk factors. Even major medical conditions—hypertension, obesity, a prior preterm birth—still can’t predict more often than not if a baby will arrive prematurely. More recently, researchers have pointed to environmental clues, like levels of air pollution or proximity to coal power plants and fracking sites; behaviors like drug and alcohol use or smoking; and social traumas, like domestic violence. But none of those alone can identify women most at risk.

In Kentucky, Passport Health, a nonprofit Medicaid insurer, is testing an unlikely hypothesis: Whether artificial intelligence can make sense of the cacophony of risk factors and direct expecting mothers into more personalized care. Last year, a Louisville startup, Lucina Health, started mining the health plan’s records for signs of women who are most at risk of an early delivery. It’s a test of whether insurers, normally associated with cost-cutting, can help tackle a national health problem.

“There is no pill [to prevent] preterm birth,” says Laura Jelliffe, who directs precision health research at the UCSF California Preterm Birth Initiative. For most women, the best interventions extend beyond the doctor’s office: a plan to manage stress or address stability at home, or help finding treatment for smoking or drug use. Insurers are an overlooked part of that equation, she adds. In theory, they have the ability to coordinate care across a patchwork of health and welfare providers, which doctors often don’t. They also have a financial incentive. An on-time birth that goes smoothly runs, on average, about $10,000; a preemie who winds up in a neonatal intensive care unit might cost more than $1 million—not including the costs of complications over a lifetime.

Insurers like Passport, however, aren’t naturally built for the job. The problem starts with an antique claims system, says Larry Griffin, an obstetrician-gynecologist and medical director for women’s health at Passport. Even identifying which plan members are pregnant is challenging. In the past, Passport simply relied on obstetricians to give them a heads up about a new pregnancy, but that only revealed about a fifth of women in their first trimester. Some weren’t seeing doctors until much later, or their insurance forms idled in the system. Others slip through the cracks entirely and only become known to the plan after they’ve delivered.

“The truth is that moms are in the system, they just can’t find them,” says Kevin Bramer, CEO of Lucina Health. “You have big, chaotic systems that fail to talk to each other.” The company starts by rifling through insurance claims—flagging, say, the set of codes for a woman who went to the emergency room with a broken arm and left with a positive pregnancy test, but hasn’t made it to an obstetrician. By drilling down into patient records and claims, and vetting against other data like prescriptions, Lucina says it’s been able to more than triple the identification rate in the first trimester to about 68 percent. Once a woman is confirmed pregnant, data from her health records (say, a history of hypertension or a prior preterm birth) and demographics (older moms are more likely to deliver preterm) are used to generate a preliminary risk score; if she rates highly, she’s assigned a nurse to learn more.

That early intervention is critical, doctors say. The only clinical test for preterm birth is a measure of cervical length, taken during an ultrasound (shorter usually indicates greater risk). But its predictive power wanes after the first trimester. Better methods are on the way, but none is poised to become widely available soon. A test being developed by Stanford researchers, for example, makes a prediction using RNA circulating in the blood, looking for genetic signs that match those found in women who gave birth early. Jelliffe is working on another test at UCSF that uses biomarkers for inflammation and immune response, but she says her team is still working on how best to commercialize the tech.

Once a set of risk factors is identified, a few situations offer clear steps: A woman who’s had a prior preterm birth might go on progesterone, for example, a hormone that helps the womb grow and prevents contractions. Early signs of hypertension—a precursor to a deadly condition called preeclampsia, which can starve the mom’s organs or the fetus of blood—leads to aspirin and a management plan. But the vast majority of women who give birth early don’t fall into those categories, Jelliffe says. Having exhausted their standard medical options, some doctors want to avoid worrying patients (chronic stress is another risk factor). As one maternal-fetal medicine specialist put it, “If there’s not anything that we can do differently, why even mention the risks?”

That’s where an insurer can step in, Passport’s Griffin says. “We’ve only recently started to look at the social determinants of health,” he says. “It’s not enough to look at hypertension and smoking. We’ve got to look at the home situation and correct food and transportation and childcare.” For the women Lucina determines to be at high risk, a nurse care manager checks in periodically, coordinating prenatal visits and making referrals for services like transportation to appointments or substance-abuse treatment. The algorithm churns daily, updating with new information from both health records and interviews. If new information emerges—a doctor’s visit reports spiking blood pressure, or an interview reveals trouble at home—the risk score gets an update, potentially bumping those with lower scores into case management.

Initial data indicates that, a year into the program, preterm births fell by about 13 percent. But Passport is still studying what that means for its bottom line. The company is spending more on the additional services for at-risk pregnancies, but it’s also saving money from fewer emergency room visits, Griffin says. Other potential benefits that still need to be measured, he adds. If a woman is aware of the signs of preterm birth, she may be more likely to seek help, leading to interventions that can help delay delivery, like progesterone or a small surgical procedure to help keep the cervix closed. Even giving the baby a few extra days to develop in the womb can keep her out of the NICU and reduce long-term complications. Lucina’s Bramer says the company will now focus on improving its predictions using outcomes from the current group of newborns, and tracking their health through their first years of life. Those insights will also help calibrate a new program being set up with a Medicaid provider in Florida.

To see bigger improvements, Griffin says he’d like to engage providers more. Passport faces its own hurdles; the health plan recently sued the state over Medicaid reimbursement cuts, which it says put the nonprofit at risk of insolvency. And Medicaid has its limitations. Kentucky, for example, is among the 47 states that don’t cover doula services for Medicaid patients. So while Passport and Pickett’s doulas may serve the same population of Medicaid-eligible women in Louisville’s West End, for now, they’re disconnected. Pickett relies on indirect referrals from other community health programs, and word-of-mouth, to see Passport patients.

Relying too much on self-reported data also carries risks. Lucina’s predictive algorithms are only as good as the data that’s gathered through interviews and checkups. Women of color, who are far likelier to report stressful and racist interactions with medical professionals, may be less willing to share intimate personal details. If some women don’t share relevant details about their lives, the algorithms won’t effectively identify those most at risk—and could even exacerbate the inequities the program is meant to close.

Griffin hopes that having nurses build a relationship over the course of a pregnancy will help. He says the data crunching is just a catalyst for getting more women in the door and into care. With about 10,000 deliveries each year among Passport’s Medicaid patients in Kentucky, risk assessment lets nurses focus on a smaller number of high-risk women, rather than spread themselves across many patients. “It takes time to break barriers down. If we don’t find out about an abusive husband on the first phone call, we might get it on the third or fourth,” he says. “It’s a change for us. We never even asked that question a decade or two ago: Do you feel safe at home?”