A large multisite cohort study published in JAMA Network Open suggests that a machine learning model using routinely collected electronic health record data may help predict short-term risk of preeclampsia in late pregnancy.
Preeclampsia, a hypertensive disorder affecting 2% to 8% of pregnancies worldwide, remains a leading cause of maternal and perinatal morbidity and mortality. Its unpredictable onset and rapid progression have limited the effectiveness of existing prediction tools, which often rely on specialized biomarkers or static risk assessments.
In this retrospective study of 58,839 pregnancies across three NewYork-Presbyterian hospitals, researchers developed and validated machine learning models to estimate the likelihood of preeclampsia onset within one, two, and four weeks. The models incorporated routinely available clinical variables, including blood pressure, laboratory results, and maternal characteristics.
Model performance improved through the third trimester, with strongest discrimination observed at approximately 34 weeks’ gestation. Predictive accuracy declined slightly near 38 weeks and increased again closer to delivery. Overall, the models demonstrated good discrimination in both internal and external validation cohorts.
The researchers reported that preeclampsia risk in late gestation can be dynamically estimated using routinely available clinical data. Unlike static models that assess risk at a single time point, this approach updates risk continuously as new information becomes available.
The models maintained high negative predictive values, indicating a strong ability to rule out near-term risk. Positive predictive values were lower, reflecting the relatively low incidence of preeclampsia, but improved as delivery approached and were higher than those of traditional risk-based models.
Blood pressure was the most influential predictor across all time points. Laboratory values contributed more earlier in the third trimester, while demographic and obstetric factors became more informative closer to term.
Performance remained consistent across the three hospitals, although all were within a single health system. Model performance improved further when tailored to site-specific data.
“These results suggest that this approach may offer opportunities for earlier intervention and could be adaptable across clinical settings,” the researchers wrote.
The authors noted that the model may provide actionable lead time for closer monitoring or decisions about delivery. However, they emphasized that the findings reflect retrospective analysis and require prospective validation before clinical implementation.
The study was limited by its retrospective design and reliance on data from a single health system.
Four researchers reported a pending patent related to integrated point-of-care testing and artificial intelligence–powered adverse pregnancy outcome technologies; no other disclosures were reported.
Source: JAMA Network Open
