Utilizing Data-Driven Insights to Improve Maternal Health

Preeclampsia is a serious condition that affects some women during pregnancy or in the postpartum period. It is diagnosed by persistent high blood pressure, often accompanied by high protein levels in the urine. The syndrome may lead to severe complications for the mother, including stroke and organ failure, and her child may experience low birth weight and even stillbirth.

Around 1 in 25 pregnancies in the United States are affected by preeclampsia, which is the second leading cause of maternal mortality worldwide. According to the Preeclampsia Foundation, there has been a 25% increase in incidence in the last two decades. A major unanswered question is how to identify which women are at risk.

Dr. Li Li, Sema4’s Vice President of Clinical Informatics, believes that the answer to that question is data – and lots of it. We interviewed Dr. Li about her research, in collaboration with a team at the Mount Sinai Health System, on using predictive models to help identify risk factors for preeclampsia. Read on to learn about how she builds these models and her findings so far.

Dr. Li Li
Dr. Li Li, VP, Clinical Informatics

Q: Why did you choose to investigate preeclampsia?

There is currently no real gold standard to determine which patients are at risk for the condition. Preeclampsia has historically been detected at one time point when high blood pressure or urine protein are measured. However, there is a wide range of data from the pre-pregnancy, intrapartum, and post-delivery periods that may guide us towards earlier detection.

Q: What are you doing to identify risk factors for preeclampsia?

Sema4 has extensive advanced informatics capabilities. We can analyze large amounts of patient data longitudinally, across the entire time series of pregnancy, from pre-pregnancy information (demographics, medications, and comorbidities) to individual visits during pregnancy (blood pressure, urine protein markers, and lab work) and the postpartum period. The goal of our work is to use this big data to build models to assess the probability of developing preeclampsia at each time point in the dynamic trajectory of pregnancy.

Q: How many patients did you study?

By analyzing electronic health records (EHRs), we retrospectively identified more than 100,000 completed pregnancy-delivery journeys. Our research spans the eight months prior to conception, through the full pregnancy journey to 10 weeks postpartum. We used deidentified data from these patients’ EHRs to create network models to predict risk for preeclampsia. We then tested these models to see how well they predicted outcomes in those retrospective journeys. Our goal now is to take the models and apply them to prospective data in a digital trial to identify women at risk before they develop symptoms.

Q: Can you tell us about any of your findings?

We built 19 models, covering time points from week four of pregnancy through postpartum, and found around 40 unique features at each time point. One particularly interesting feature was urinary protein; high levels are associated with preeclampsia, but traditionally, doctors might ignore a trace amount. However, we found that trace amounts were present in more than 50% of the patients who went on to develop preeclampsia.

A second noteworthy finding is that many markers change over time. So, in the preeclampsia population, a signal that is not detectable at the start of pregnancy might be high at the end or vice-versa. There are preeclampsia-specific patterns that we can spot.

Finally, another interesting marker is systolic blood pressure. A repeated systolic blood pressure of 140, in combination with urinary protein, is considered the definition of preeclampsia. Our analysis showed that at the antepartum, intrapartum, and postpartum stages, a reading of 130 is already associated with a higher relative risk of developing preeclampsia.

Q: How will your findings influence maternal care and treatments?

Our end goal is to apply our findings to clinical care, in combination with other predictive models. Multiple models could be integrated into a resource to assess disease comorbidities and risk at patient visits to improve patient care. There is currently no prevention or good treatment for preeclampsia, so the aim is to give the physician an early warning. This advance notice can then lead to enhanced management and monitoring of a potential high-risk patient.

Our research also showed that ibuprofen offers a protective effect to lower the risk of preeclampsia in the postpartum period. This result could have potential clinical utility.

Q: What other aspects of pregnancy will you investigate next in your dataset?

We have already developed predictive models for postpartum hemorrhage and gestational diabetes. Postpartum hemorrhage is another serious pregnancy complication with few warning signs. We identified at least 24 predictive markers for it that are already routinely measured in clinical care. Again, if we can alert a physician to a high-risk case, they can prepare more fully and provide better patient care.

_ _ _

Want to know more about Dr. Li’s preeclampsia research? Check out her latest preprint paper on medRxiv here.