Graduate Center professor uses AI to weigh the risks of preeclampsia

June 22, 2022

Team designs first machine-learning system to predict severe preeclampsia.

Professor Anita Raja
Professor Anita Raja co-led a study that used artificial intelligence to predict severe preeclampsia, a potentially life-threatening type of hypertension in pregnant women. (Photo credit: Alex Irklievski)

Each year, about 700 women die from pregnancy-related complications in the United States. One of the leading causes of maternal death is preeclampsia, a type of hypertension that develops midway through pregnancy. If left untreated, the condition can be fatal to both mother and baby.

Symptoms include high blood pressure, headaches, swelling in the hands and face, blurred vision, abdominal pain, nausea, and vomiting. There may even be no symptoms. Serious complications include preterm birth, liver and kidney damage, and even seizures. The disease can progress quickly, so early detection is crucial. 

In a soon-to-be-published study, Professor Anita Raja (GC / Hunter, Computer Science) and a team of researchers report on the use of artificial intelligence to predict severe preeclampsia in first-time mothers. The project is funded by the National Institutes of Health, which named the project a winner of its Decoding Maternal Morbidity Challenge in February. 

A reliable indicator 

The goal of the challenge was to come up with new ways to use data previously gathered from thousands of pregnant women to assess maternal health outcomes. The group used computer modeling and artificial intelligence to predict severe preeclampsia in first-time mothers. 

They found that by looking closely at chemical compounds found in the placenta, or placental analytes, they could predict the early onset of severe preeclampsia. “There were specific analytes that we studied, which showed that for those patients, later on, preeclampsia was a given,” said Raja, one of the study’s co-authors and director of the Distributed Artificial Intelligence Research lab at Hunter College.

Learn more about the Ph.D. program in Computer Science

The scientists built a machine-learning system with existing data gathered from a previous NIH study involving 10,000 pregnant, first-time mothers. They started by first looking at the known risk factors of preeclampsia. 

“Generally, the indicators that are known are high blood pressure and proteinuria,” Raja said, using the medical term for elevated protein in the urine. Other indicators include cerebral and visual problems, liver dysfunction, and certain blood conditions, she said.

But when the data was examined closer, the team saw something that piqued their interest. “Two things really stood out — blood pressure and placental analytes,” she said. 

These analytes included pregnancy-associated plasma protein and placental growth factor (PIGF) — a protein molecule critical to the healthy formation of the placenta — taken from maternal blood samples. “We saw that decreased levels of PIGF, during the first and second trimesters, precede the onset of preeclampsia,” Raja said. 

Additionally, the group found that traditional risk factors such as age, race, sleep apnea, and history of preeclampsia were not significant predictors of severe preeclampsia. “We observed that these factors, which are commonly known in the literature, didn't show up, at least not in our study,” Raja explained. In line with existing research, however, the data showed high blood pressure to be a key risk factor, particularly in the third trimester, Raja said.

Pregnancy is a dynamic state

The team applied machine-learning algorithms to four different models. “Pregnancy is a dynamic state,” said Raja. “We can't just build one model covering nine months. What we did was split it up into four different machine-learning models. We looked at, ‘What are the factors for the first six to 13 weeks,’ and then 16 to 21 weeks, and so on.” 

Machine learning begins with existing data, Raja said. “So, we know what the results are for those patients,” she said. “And when a new patient comes in, you use that model and compare and say, ‘What is the risk of this patient? How did they compare to those patients we have already have in our training set? Are they similar to the ones who are high risk or are they similar to the ones who are low risk?’”

They found that BMI was a serious risk factor during the first visit, compared to visits later on. “The model for visit three indicated that blood pressure had a higher impact. So, overall, blood pressure is an important factor, and placental analytes is what stood out for us across all models.”

The takeaway was that placental analytes should be collected in patients who are at risk of developing preeclampsia, Raja said. “Placental analytes are powerful indicators to early onset of severe preeclampsia,” Raja said. “It was a very important observation from our perspective.” 

Bias-free models

Their goal was to build models that were bias-free as possible, Raja said. “Initially, when we looked at our models, it seemed like non-Hispanic, Black women were being predicted to have a higher risk of preeclampsia, in general,” said Raja. “This happened because it was not equal representation of white and Black patients, and for Asians, it was even smaller,” Raja said. 

The issue came down to math. About a fourth of patients from the original group had been tested for placental analytes. And, after other exclusions, the team was left with a final cohort of less than 1,800 mothers. Of this subgroup, only a small number of Black patients remained.   

“We had to correct our model,” she said. “Then we found that race was not really an issue, which was surprising but important. We don’t want to give unnecessary treatment or a lack of treatment because that could lead to loss of the baby or unnecessary health issues.”

The research will provide needed data to help predict and prevent cases of severe preeclampsia, Raja said. The findings are expected to be published later this year. “It's given us a better understanding of the disease,” she said. “We can find out who is at risk, and for those who are at risk we might do extra data collection to measure analytes from maternal serum.”

For Raja, the results are clear. “We are now saying that placental analytes are important,” she said. “And that's a big deal.” The findings will contribute to data from previous studies that also found placental analytes to be good predictors of preeclampsia, Raja added.

The team includes researchers from the Graduate Center, Hunter College, Columbia University Irving Medical Center, the Computer Science Department at Columbia University, and the Indiana University School of Medicine. The project was funded through the National Institute of Child Health and Human Development (NICHD). 

Published by the Office of Communications and Marketing