Machine learning model predicts preeclampsia risk week by week in the third trimester

Most preeclampsia cases arise in the third trimester. Most predictive models are designed for the first. That mismatch leaves clinicians without a reliable tool precisely when they need one most.

A team at Weill Cornell Medicine has developed a machine learning model that works in the opposite direction: it takes electronic health record data from late pregnancy and provides continuously updated risk estimates as the third trimester unfolds. Published March 6, 2026, in JAMA Network Open, the study describes a model trained on nearly 36,000 pregnancies and validated at two additional hospitals.

The gap in existing tools

Preeclampsia, a sudden-onset condition involving high blood pressure before delivery, affects 2% to 8% of pregnancies worldwide. It can cause seizures, organ damage, and death for both parent and child. First-trimester screening models exist and are useful for initiating preventive measures like aspirin early in pregnancy. But these tools were designed to catch early-onset preeclampsia, which accounts for a minority of cases. Late-onset and term preeclampsia, which make up the majority, are poorly predicted by early screening.

The Weill Cornell team, co-led by Dr. Fei Wang and Dr. Zhen Zhao with clinical expertise from Dr. Tracy Grossman, set out to fill this gap with a model that operates during the weeks when preeclampsia is most likely to develop.

Training on 36,000 pregnancies

The researchers built their model using deidentified electronic health record data from 35,895 pregnancies at NewYork-Presbyterian/Weill Cornell Medical Center between October 2020 and May 2025. Rather than producing a single risk score at one point in time, the model continuously recalculates preeclampsia probability as new clinical measurements are recorded throughout the third trimester.

The model performed best around 34 weeks of gestation, potentially providing clinicians with several weeks of lead time before delivery to adjust monitoring and management.

Different predictors at different stages

One of the study's more interesting findings is that the most important predictive variables change as the third trimester progresses. Blood pressure was the strongest predictor throughout, which is expected given that hypertension defines preeclampsia. But the supporting cast shifted.

Early in the third trimester, abnormal routine blood test results emerged as important secondary predictors. The researchers suggest these may reflect emerging problems with the placenta, the organ that supplies nutrients and oxygen to the fetus. Placental dysfunction is a known contributor to preeclampsia, and laboratory abnormalities may signal it before blood pressure rises dramatically.

Later in the third trimester, the patient's age and white blood cell count became more prominent predictors, suggesting that systemic inflammation may play a larger role in preeclampsia that develops closest to term. If these stage-specific patterns are confirmed, they could point toward different underlying mechanisms at different gestational ages, a hypothesis that has been debated in the field.

Validation across three hospitals

The model was validated on data from 8,664 pregnancies at NewYork-Presbyterian Lower Manhattan Hospital and 14,280 at NewYork-Presbyterian Brooklyn Methodist Hospital. These hospitals serve different patient populations than the training site, providing a test of whether the model generalizes across demographics and practice patterns.

What it does not do

The model predicts risk; it does not prevent preeclampsia. Whether acting on its predictions improves outcomes depends on what clinicians do with the information. Enhanced monitoring, blood pressure management, and decisions about delivery timing are all reasonable responses to elevated risk, but none have been tested in the context of this specific model.

The study also cannot determine whether preeclampsia at different points in the third trimester has genuinely distinct causes. The shifting predictor importance is suggestive but not conclusive. Prospective studies designed to test stage-specific interventions would be needed to confirm the clinical relevance of these patterns.

The model relies on electronic health record data, which means its accuracy depends on the quality and completeness of routine clinical documentation. It has not been tested in healthcare systems outside the NewYork-Presbyterian network, and differences in documentation practices, patient populations, and clinical protocols could affect performance.