AI preeclampsia prediction can identify high-risk pregnancies early, but prediction only helps when it leads to action. The proven action is low-dose aspirin started before 16 weeks, plus closer blood pressure monitoring. Where women book late, aspirin is unavailable or follow-up is thin, even an accurate model changes little.
Every year, hundreds of thousands of women die from complications of pregnancy and childbirth. The World Health Organization estimates about 260,000 maternal deaths in 2023, around 92% of them in low- and lower-middle-income countries and about 70% in sub-Saharan Africa. High blood pressure disorders, preeclampsia and eclampsia among them, are one of the five complications that together cause about three-quarters of those deaths. Preeclampsia also drives preterm birth, because the only cure is delivering the placenta, often early.
That makes prediction look like an obvious target for artificial intelligence. Pregnancy generates exactly the kind of repeated, structured data that models thrive on: blood pressures, weights, ultrasound measurements, blood tests, obstetric history. Startups and academic groups have built models using electronic health records, placental biomarkers and even circulating RNA. Some of the results are striking.
But the thesis here is less about the models than about what surrounds them. Early risk prediction for preeclampsia matters because we have one cheap, proven preventive step and a window in which to take it. A model’s value is almost entirely determined by whether a health system can act inside that window. In many of the places where preeclampsia kills most, it cannot yet, and fixing that is a bigger lever than a better algorithm.
Can AI predict preeclampsia before symptoms appear?
Yes, to a meaningful degree, and the best-established tool is not what most people mean by AI.
The Fetal Medicine Foundation’s first-trimester screening uses a statistical model, based on Bayes’ theorem and competing risks, that combines maternal characteristics and history with mean arterial blood pressure, uterine artery Doppler on ultrasound and a blood marker called placental growth factor. In the SPREE study published in Ultrasound in Obstetrics & Gynecology in 2018, Tan and colleagues compared this approach with the UK’s then-standard NICE checklist of risk factors in 16,747 pregnancies. The NICE method detected 40.8% of women who went on to develop preterm preeclampsia, at a screen-positive rate of about 10%. The full Bayesian model detected 82.4%.
That is a doubling of detection from better use of data, and it came from careful statistical modelling rather than deep learning. It also relies on inputs, uterine artery Doppler and placental growth factor, that require trained sonographers and laboratory capacity many health systems do not have.
Newer approaches push further. In 2022, Rasmussen and colleagues reported in Nature that signatures in cell-free RNA from a single maternal blood sample could predict preeclampsia, with 75% sensitivity and a positive predictive value of about 32%, on average about 14.5 weeks before delivery. Other groups have trained machine learning models on routine electronic health record data, aiming to flag risk from information already collected at booking.
Two cautions apply to all of them. Retrospective accuracy in the population a model was built on often falls when it moves to a new population with different ethnicity, age, parity and health service patterns. And a positive predictive value around 30% means most women flagged will not develop the condition, which matters when the flag triggers extra visits and anxiety.
What happens after a woman is flagged as high risk?
This is where the evidence becomes unusually clear.
The ASPRE trial, published in the New England Journal of Medicine in 2017 by Rolnik and colleagues, took women identified as high risk by the first-trimester screening model and randomised them to aspirin 150 mg a day or placebo, from 11 to 14 weeks until 36 weeks. Preterm preeclampsia occurred in 1.6% of the aspirin group and 4.3% of the placebo group, an odds ratio of 0.38. Aspirin is among the cheapest drugs in any formulary.
Screening plus aspirin is therefore one of the few prediction-to-prevention chains in obstetrics with randomised evidence behind the whole sequence. National and professional guidance has moved accordingly. In the United States, the Preventive Services Task Force recommends low-dose aspirin after 12 weeks for people at high risk of preeclampsia, defined mainly by clinical risk factors. The WHO recommends low-dose aspirin for women at high risk as part of antenatal care.
But notice the conditions built into that chain. The woman has to be seen in the first trimester. Her risk has to be assessed then, not at 28 weeks. Aspirin has to be available, affordable and actually taken daily for months. And after that, someone has to keep measuring her blood pressure and checking for protein in her urine, because aspirin reduces risk rather than eliminating it, and preeclampsia that does develop needs timely recognition, antihypertensive treatment, magnesium sulphate for severe disease and a plan for delivery.
Break any link, and the value of the prediction falls toward zero.
A worked example: two districts, one model
Hypothetical, with round numbers for illustration only. Imagine the same prediction model deployed in two districts, each with 10,000 births a year and a preterm preeclampsia rate of 2%, so 200 affected pregnancies.
In District A, 85% of women attend antenatal care before 14 weeks, aspirin is stocked at every clinic, and there is a working referral system for hypertension. The model flags 10% of women, catching 75% of those who will develop preterm preeclampsia. Of the 200 women who would be affected, about 128 are both seen early and flagged. If aspirin roughly halves their risk, as trial evidence suggests is plausible, dozens of cases are prevented, and those that occur are caught early.
In District B, only 30% of women attend before 14 weeks, aspirin stockouts are frequent, and hypertension referral depends on transport that is unreliable. The same model, with the same accuracy, reaches about 45 of the 200 at-risk women in time. If only half of those get and take aspirin, a much smaller number of cases is prevented. The women the model could not reach, because they booked late, are exactly the ones most likely to die of undetected disease.
The algorithm is identical. The benefit differs several-fold, and the difference comes entirely from early booking, drug supply and follow-up. A health system with District B’s profile would likely save more lives by getting women into antenatal care earlier and making blood pressure checks reliable than by buying a better model. This is the same lesson we drew from cancer screening in the screening gap is about the second visit, not the test.
What about AI to predict preterm birth?
Preterm birth prediction is harder, partly because preterm birth has many causes. Short cervical length on ultrasound in mid-pregnancy is a known risk marker, and progesterone treatment or cervical cerclage can help selected women. Several blood tests and machine learning models aim to predict spontaneous preterm birth earlier or more accurately.
The same logic applies with more force. Prediction is valuable when it leads to an intervention that works for the predicted group, delivered in time. For preterm birth, the effective interventions are narrower and more dependent on specialist services than aspirin for preeclampsia. A model that identifies high risk in a setting with no cervical length scanning, no progesterone and no neonatal unit able to care for a baby born at 30 weeks produces worry rather than benefit. Where a model’s greatest value may lie is in routing women to facilities with neonatal care before labour starts, which is an operational rather than a diagnostic use.
Is the case for AI prediction stronger than this?
The strongest counterargument is that AI can reduce dependence on the very resources that limit current screening. If a model built on routine data such as age, blood pressure, weight, parity and history can approach the accuracy of screening that needs Doppler ultrasound and biomarkers, it could bring first-trimester risk assessment to places that will never have specialised sonographers. Smartphone-based blood pressure tools and simple point-of-care tests could extend this further.
There is also a case that even in settings with late booking, prediction later in pregnancy is useful: a high-risk flag at 24 weeks may not allow aspirin to work as intended, but it can prompt more frequent blood pressure checks and earlier referral, which matter for preventing deaths from eclampsia and stroke.
Both points are reasonable, and the first is a genuine research priority. But both reinforce rather than undermine the thesis. The model is useful when it is designed around what the system can do next, and the actions, aspirin, blood pressure checks, referral, magnesium sulphate, are what save lives. For more on why digital tools in low-resource settings stall without people and processes behind them, see our piece on digital health in low-bandwidth settings.
What this means for health systems, clinicians and developers
For ministries and health systems, the priority order is clear: early antenatal booking, reliable aspirin and antihypertensive supply, routine blood pressure measurement at every visit, and working referral for severe disease. A prediction model sits on top of that foundation, and it should be evaluated by how many at-risk women it gets onto aspirin in time.
For clinicians, a model’s output is a risk estimate, not a diagnosis. The response to a high-risk flag should follow established guidance on aspirin and monitoring, and women at low predicted risk still need their blood pressure checked, because some of them will develop the disease.
For developers, build for the inputs the target system actually has, validate in the population where the tool will be used, report positive predictive value and the screen-positive rate as well as sensitivity, and measure what happens after the flag, including aspirin uptake and adherence.
For funders, prediction research is valuable, but money spent on supply chains and early booking may produce more benefit per dollar where those are weak. The two are complements, not substitutes.
For pregnant women, the practical point is simple: attend antenatal care early, ask whether you should take low-dose aspirin, and discuss any headache, vision changes or swelling promptly with your midwife or doctor.
Frequently asked questions
Can AI predict preeclampsia?
Models can identify many women at high risk early in pregnancy. The best-validated first-trimester screening, a statistical model combining history, blood pressure, Doppler ultrasound and placental growth factor, detected 82.4% of preterm preeclampsia in a 2018 UK study, versus 40.8% for a risk-factor checklist. Machine learning and blood-based RNA tests show promise but need validation in diverse populations.
Does aspirin prevent preeclampsia?
In women identified as high risk, yes. The ASPRE trial in the New England Journal of Medicine (2017) found aspirin 150 mg daily from 11 to 14 weeks until 36 weeks reduced preterm preeclampsia from 4.3% to 1.6%. Guidance in many countries recommends low-dose aspirin for high-risk women. Women should discuss whether aspirin is right for them with their clinician.
Why does early antenatal booking matter for preeclampsia prediction?
The strongest preventive evidence involves aspirin started before about 16 weeks. If a woman’s first antenatal visit is later, the best window for prevention may be missed, however accurate the risk model. Earlier booking also allows baseline blood pressure measurement, which makes later rises easier to spot.
Is AI preeclampsia screening useful in low-income countries?
It could be, if it uses inputs available locally and leads to action. Most maternal deaths occur in low- and lower-middle-income countries, where specialised ultrasound and lab tests are often unavailable. A model using routine data could extend risk assessment, but only with reliable aspirin supply, blood pressure monitoring and referral for severe disease.
The history of maternal health is full of good ideas that failed at the next step. AI prediction for preeclampsia is a good idea. Its success will be measured in aspirin taken on time and blood pressures checked at every visit, not in the area under a curve.
Where you work, what is the weakest link between identifying a high-risk pregnancy and acting on it?
