Model Could Sound the Alarm on Preeclampsia Risk in Late Pregnancy

A machine learning model for prediction of preeclampsia risk using routinely collected data was feasible among pregnancies in late gestation, according to a retrospective cohort study.

Among 58,839 pregnancies across three New York City hospitals, the model’s predictive performance peaked at 34 weeks’ gestation with area under the receiver operating characteristic curves of 0.863 at training and 0.808-0.834 at validation, reported Zhen Zhao, PhD, of Weill Cornell Medicine, and colleagues in JAMA Network Open.

The model’s positive predictive value was low in the early observation period, increasing from approximately 0.001 to 0.002 at 28 weeks’ to peak values at 36 weeks’ (mean 0.057 at one hospital and 0.046 at another). The negative predictive value remained high across all gestational ages (≥0.993).

“Our study shows that preeclampsia risk can be dynamically predicted using information already collected in routine prenatal care,” Zhao told MedPage Today.

Across the three hospitals, patients who developed preeclampsia were older and more frequently Black.

Preeclampsia is a serious hypertensive disorder of pregnancy and a leading cause of maternal and perinatal morbidity and mortality. It affects 2% to 8% of pregnancies worldwide.

“[Preeclampsia’s] unpredictable onset and rapid progression pose critical challenges for obstetric care, making timely risk prediction a major unmet need,” the authors wrote.

However, Zhao noted, most existing tools focus on risk prediction in early pregnancy or rely on specialized biomarkers or imaging tests. For this study, the authors took a different approach by utilizing routinely collected electronic health record data. The machine learning model’s inputs included demographic and obstetric characteristics, blood pressure, and lab test results. Self-reported race data were also collected.

Zhao and team found that the most informative predictor was blood pressure. Measures like albumin, alkaline phosphatase, and hematologic indexes contributed to earlier gestation, while demographic and obstetric factors were of increasing importance in later pregnancy.

“This shows that routine laboratory data provide additional predictive value beyond blood pressure alone, and these data are already collected as part of standard clinical care,” Zhao said.

David Hackney, MD, of Case Western Reserve University in Cleveland, who was not involved in the study, told MedPage Today that the authors conducted a highly technical analysis of a robust dataset. His main question, he noted, is what to do with predictions of preeclampsia risk.

Prophylactic aspirin administration is increasingly universal, and pharmacologic normalization of blood pressure elevations in the first half of pregnancy is also becoming more widespread.

“Outside of blood pressure control and aspirin prophylaxis, we do not have many strategies to mitigate preeclampsia onset and the only cure remains delivery,” Hackney said. “Thus, if I was handed a magical crystal ball which would tell me exactly when and if each patient would experience preeclampsia, it’s not clear how I would use it beyond either providing the patient with reassurance or allowing them to emotionally and logistically prepare.”

Hackney added that if prediction models like this are implemented, it’s important to consider whether race should be included as a clinical variable.

“As race is often a flawed proxy, the real-world implementation of predictive models that are inclusive of racial predictors risks disparity exacerbation,” he pointed out.

For this study, Zhao and team included pregnancies that delivered from October 2020 through May 2025 at three NewYork-Presbyterian hospitals. The model was trained at Weill Cornell Medical College (WCMC) and validated at Lower Manhattan Hospital (LMH) and Brooklyn Methodist Hospital (BMH). The primary outcome was development of preeclampsia within specified prediction windows, which was identified via ICD-10 code.

Mean maternal age was 33.3. Most patients were white (55.5%), 17.3% were Asian, 11.1% were Black or African American, and 16% were other races.

Across all sites, preeclampsia pregnancies were characterized by older maternal age (median age 35 vs 34 in the WCMC group, P<0.001; 35 vs 34 in the LMH group, P=0.003; and 33 vs 31 in the BMH group, P<0.001), and greater proportions of Black race (15% vs 6.5% in the WCMC group, P<0.001; 14.8% vs 7.2% in the LMH group, P<0.001; and 41.8% vs 21.8% in the BMH group, P<0.001).

Nulliparity, multifetal gestation, and higher pregravid body mass index were also higher in preeclampsia pregnancies compared with controls.

The study was limited by its retrospective, single-health system design; by the potential for misdiagnosis; and missing data and irregular sampling in the large longitudinal dataset.

Zhao noted that in the future, this model could be useful in resource-limited settings since it relies primarily on blood pressure and routine laboratory data.