The model continuously recalculates risk as new clinical information arrives
Among the most dangerous surprises in obstetric medicine, preeclampsia has long resisted prediction in the final weeks of pregnancy — the very window when intervention matters most. Researchers at Weill Cornell Medicine have now built a machine-learning model that watches continuously, recalculating risk as each new vital sign and lab result arrives, offering clinicians something rare in late pregnancy: time to act. Trained on tens of thousands of pregnancies and validated across multiple hospital cohorts, the system points toward a future where a condition that has long announced itself too late may finally be seen coming.
- Preeclampsia strikes 2–8% of pregnancies worldwide and can escalate to organ failure, seizures, or death for parent or child — often before clinicians have adequate warning.
- Unlike earlier tools that issued a single early-pregnancy risk estimate, this model continuously updates predictions throughout the third trimester as real-time clinical data flows in.
- The system performed most powerfully at 34 weeks of gestation, giving clinicians a meaningful lead time before delivery — the precise window where intervention can shift outcomes.
- Blood pressure proved the dominant predictor, but the model also revealed a shifting constellation of risk signals: placental markers early, inflammatory indicators like white blood cell counts closer to delivery.
- The research, published in JAMA Network Open, raises the possibility that preeclampsia is not one disease but several, each with its own biological signature — a finding that could eventually guide more targeted treatment.
Preeclampsia strikes between 2 and 8 percent of pregnancies worldwide, and when it arrives in the third trimester — when time is shortest and stakes are highest — it has long been among obstetrics' most dangerous surprises. A team at Weill Cornell Medicine in New York has built a machine-learning system designed to change that.
Unlike earlier screening tools that offered a single risk estimate and moved on, this model continuously recalculates a patient's preeclampsia risk as new clinical information arrives throughout the third trimester. Trained on nearly 36,000 pregnancies at NewYork-Presbyterian/Weill Cornell Medical Center and validated against more than 23,000 pregnancies at two other hospitals, the system showed its strongest predictive performance at around 34 weeks of gestation — the point where clinicians gain the most valuable window to act before delivery.
Blood pressure emerged as the single strongest predictor, consistent with what clinicians already understand about the condition. But the model revealed something more layered: the warning signs shift as pregnancy advances. Early in the third trimester, abnormal blood markers suggesting placental dysfunction carried the most weight; later, a patient's age and white blood cell counts grew increasingly predictive, pointing toward inflammatory processes. Published in JAMA Network Open, the findings hint at what the field has long suspected — that preeclampsia may be several diseases, each with its own biological signature.
The practical stakes are immediate. Real-time risk awareness allows clinicians to tighten monitoring, manage hypertension more aggressively, and make more deliberate decisions about delivery timing. Even a few days of lead time can mean the difference between a managed crisis and a catastrophe. The researchers now hope to determine whether the distinct biological pathways they've identified might one day allow for more targeted treatment — moving the field from reaction toward anticipation.
Preeclampsia arrives without warning in the final months of pregnancy. It strikes between 2 and 8 percent of pregnant people worldwide, and when it does, the consequences can be catastrophic—organ damage, seizures, death for mother or child. Doctors have long had tools to catch it early, but the condition that emerges in the third trimester, when time is shortest and stakes are highest, has remained stubbornly difficult to predict. A team at Weill Cornell Medicine in New York has now built a machine-learning system designed to change that.
The model works differently from the screening tests that came before it. Rather than offering a single risk estimate early in pregnancy and calling the job done, this system continuously recalculates a patient's preeclampsia risk as new clinical information arrives—blood pressure readings, lab results, vital signs—throughout the third trimester. The researchers trained it on data from nearly 36,000 pregnancies at NewYork-Presbyterian/Weill Cornell Medical Center between October 2020 and May 2025, then tested it against separate cohorts of more than 23,000 pregnancies at two other NewYork-Presbyterian hospitals. The model performed best at around 34 weeks of gestation, the point where it could give clinicians the most valuable window to act before delivery.
Blood pressure emerged as the single strongest predictor of risk, a finding that aligns with what clinicians already know about the condition's mechanics. But the model revealed something more nuanced: the constellation of warning signs shifts as pregnancy advances. Early in the third trimester, abnormal blood test results—markers that might reflect placental dysfunction—carried significant weight. Later, as delivery approaches, a patient's age and white blood cell counts became increasingly predictive, suggesting that inflammatory processes may play a growing role in the condition's development at that stage. The research, published in JAMA Network Open, hints at something the field has long suspected: preeclampsia may not be a single disease but several, each with its own biological signature.
The practical implications are immediate. If a clinician knows a patient's risk is climbing in real time, they can tighten their watch—more frequent blood pressure checks, more aggressive management of hypertension, more careful consideration of when and how to deliver. That lead time, even a few days or weeks, can mean the difference between a managed crisis and a catastrophe. For the millions of pregnant people affected by this condition each year, the ability to see it coming, rather than react after it arrives, could reshape outcomes.
The researchers acknowledge that more work remains. They want to understand whether the different biological pathways they've identified—placental dysfunction, systemic inflammation—truly represent distinct forms of the disease, and whether that knowledge might eventually allow for more targeted treatment. For now, the model represents a step toward a future where late-pregnancy preeclampsia, long one of obstetrics' most dangerous surprises, becomes something clinicians can anticipate and manage.
Citações Notáveis
By continuously updating predictions using real-time clinical data, the model may help clinicians identify patients at elevated risk of preeclampsia during late pregnancy— Weill Cornell Medicine research team
A Conversa do Hearth Outra perspectiva sobre a história
Why does preeclampsia become harder to predict the later it develops? Isn't more data supposed to make prediction easier?
You'd think so, but preeclampsia in the third trimester is a different animal. Early screening catches the cases that announce themselves through consistent patterns. The late ones are quieter, more variable. They may involve different biological mechanisms entirely—placental problems versus inflammatory cascades. The model had to learn that there isn't one preeclampsia; there are several.
So blood pressure is the obvious signal. What surprised the researchers about the other predictors?
That white blood cell counts and maternal age became more important as delivery approached. It suggests inflammation is a player, especially late. Early on, you're looking at placental dysfunction. Later, you're looking at the body's response to pregnancy itself. The disease evolves.
Does this model actually change what doctors do, or does it just confirm what they already suspect?
It changes the timing and intensity. A doctor might monitor a patient every two weeks. With real-time risk updates, they might monitor every few days. That's not confirmation—that's actionable intelligence. It's the difference between watching and intervening.
What happens to the patients who get flagged as high-risk at 34 weeks?
That's where the lead time matters. A doctor might intensify blood pressure management, increase monitoring frequency, or make an informed decision to deliver early if the risk becomes acute. You're buying time to prevent a crisis.
And the patients who aren't flagged?
That's the harder question. No model is perfect. Some will still develop preeclampsia despite low predicted risk. That's why the researchers emphasize this is a tool to support clinical judgment, not replace it.