AI Prediction Models in Pregnancy: Preeclampsia, Preterm Birth, and Gestational Diabetes

In obstetrics, the most valuable information is knowing something before it happens. If we could predict preeclampsia at 11 weeks, we could start aspirin prophylaxis in the right woman at the right time; if we could detect preterm-birth risk early, we could plan antenatal corticosteroids and referral. Traditional risk scores (history, blood pressure, biomarkers) do part of this job, but in the past three years AI has entered the stage with claims of improving prediction accuracy by weighing many variables simultaneously. In this article I examine, with current evidence, where AI prediction models for preeclampsia, preterm birth, and gestational diabetes have arrived and where they require caution.

Predicting Preeclampsia: The Most-Studied Area

Preeclampsia is among the leading causes of maternal and perinatal mortality and is, accordingly, the most intensively studied obstetric topic in AI. A systematic review covering 11 studies developing and validating machine-learning models between 2020 and 2025 includes a total of 116,253 pregnancies. The studies span geographically diverse regions — the United States, South Korea, China, Romania, Mexico, Australia, and Spain — showing that AI in obstetric care has become a global undertaking.

What do the models use as inputs?

The common logic of current approaches is to combine maternal demographics, biophysical parameters (especially mean arterial pressure), and first-trimester biochemical markers to strengthen early detection. This is essentially the long-known "combined screening" logic, scaled by machine learning.

Sharpening triage: PIERS

In another application, the machine-learning-based PIERS (Pre-eclampsia Integrated Estimate of Risk) model separates women into the very low and very high short-term serious-outcome ends more cleanly than clinical judgment. The practical implication is meaningful: it can sharpen triage in decisions about which patient needs referral, which requires intensive-care preparation, and which can be safely monitored.

Predicting Preterm Birth

Preterm birth is the single largest driver of neonatal morbidity and mortality. Here AI models combine EHR data, cervical-length measurements, and biomarkers to stratify risk. Models in this area often share infrastructure with preeclampsia models; PROMPT, for example, predicts both preeclampsia and preterm outcomes under one roof. Even so, the multi-etiology nature of preterm birth (infectious, mechanical, idiopathic) makes it difficult for a single model to capture all pathways — an honest limitation worth stating.

Gestational Diabetes (GDM) Prediction

GDM, when detected late, carries risks for both mother and baby; early prediction buys time for lifestyle intervention and closer follow-up. AI-based high-risk pregnancy prediction frameworks try to model GDM with demographic and early-pregnancy data. The main clinical gain here is the possibility of prioritizing screening to high-risk groups rather than applying it uniformly to all.

Table: Notable Preeclampsia AI Models

The "High AUC" Trap: Accuracy ≠ Benefit

The most common misreading of these models is assuming that impressive AUC automatically translates to clinical benefit. Yet the current literature itself is cautious on this point. Three things deserve emphasis:

• Imbalanced data and external validation: In relatively rare outcomes such as preeclampsia, data is imbalanced; resampling and ensemble methods are needed. A model performing well at the center where it was developed does not mean it will perform the same way in another population — external validation is essential.
• Anxiety and discrimination risks: A current analysis highlights that AI-based high-risk pregnancy prediction must balance early detection against unnecessary anxiety and stigmatization/discrimination. A false-positive "high-risk" label can harm the patient.
• The intervention chain: A prediction is valuable only when followed by an effective intervention (for example, aspirin at the right time). If the prediction is correct but nothing can be done, the model only produces worry.

Clinical Perspective

As an OB/GYN, I see these models as tools that "direct my attention" rather than "make decisions." PROMPT's retinal approach or PIERS's triage sharpness are genuinely exciting; but when I apply them to my own patients, I always ask the same question: Was this model validated in my population, and does the information it gives change a decision I can act on? AI is a candidate for one of obstetrics' strongest early-warning systems in areas such as preeclampsia and preterm birth — provided we evaluate it not by accuracy numbers but by the concrete benefit it adds for the patient.