Artificial Intelligence and Risk Models in Preeclampsia Prediction

Preeclampsia is a multisystem hypertensive disorder that affects approximately 2-5% of pregnancies and is a leading cause of maternal-perinatal mortality and morbidity. The clinical critical point is that early (especially preterm) preeclampsia can be significantly reduced by an effective pharmacological measure—low-dose aspirin started early in pregnancy. This takes “prediction” from an academic exercise to a direct matter of actionable screening: if we can identify the high-risk pregnant woman in the first trimester, we can initiate preventive treatment in a timely manner.

This is where predictive models and, increasingly, machine learning (ML — statistical/algorithmic methods that automatically learn patterns in data) come into play. In this article, starting from the competing-risks model currently used in clinical practice, we will look at newer ML/AI-based models and their limits to external validation (testing whether the model maintains the same performance in pregnant women outside the population for which it was developed), which must be honestly addressed.

First trimester combined screening: basic logic

The framework of modern preeclampsia screening is combined screening performed at 11–13⁺⁶ weeks of pregnancy. Here, “combined” means combining several independent sources of information rather than relying on a single marker:

• Maternal factors: a priori risk determinants such as age, body mass index, ethnicity, chronic hypertension/diabetes, history of preeclampsia in a previous pregnancy, method of achieving pregnancy.
• MAP (mean arterial pressure — standardized measured blood pressure value).
• UtA-PI (uterine artery pulsatility index — parameter reflecting placental blood supply resistance measured by Doppler ultrasonography).
• Biomarkers: serum PlGF (placental growth factor—an angiogenic protein that decreases in preeclampsia) and/or PAPP-A (pregnancy-associated plasma protein-A).

The basic premise of this approach is that the deviation of biomarkers from normal in preeclampsia is proportional to how early and severe the disease will be: in early (preterm) preeclampsia, MAP and UtA-PI values ​​increase significantly, while PlGF and PAPP-A decrease. Therefore, the performance of screening is strongest in early/preterm disease.

Fetal Medicine Foundation (FMF) competing-risks model

The reference framework in this field is the competing-risks model developed by the London-based Fetal Medicine Foundation (FMF). The method relies on classical statistics rather than artificial intelligence: prior risk from maternal factors is combined with multiples-of-the-median (MoM) transformations of biomarker values via Bayes' theorem to calculate an individualized preeclampsia risk for each pregnant woman. The term "competing-risks" refers to a statistical framework that models gestational age at delivery on a continuous time axis rather than using a fixed binary cutoff (e.g., 37 weeks).

Key performance data for the model come from O'Gorman et al.'s prospective study of 35,948 singleton pregnancies (with Akolekar, Poon, Nicolaides; AJOG 2016). The combination of maternal factors + MAP + UtA-PI + PlGF has a fixed false positive rate (FPR) of 10%:

• 75% (95% CI 70-80) of preterm preeclampsia (<37 weeks),
• Detected 47% (95% CI 44-51) of term preeclampsia (≥37 weeks).

These values are significantly superior to the performance of screening with maternal factors alone under the same conditions (49% and 38%, respectively). There is an important and often overlooked detail here: in this study, the addition of PAPP-A did not significantly improve the performance of screening — the most valuable triplet for preterm preeclampsia is MAP, UtA-PI, and PlGF, in addition to maternal factors (the term “triple test” frequently used in the literature refers to these three biomarkers).

One of the strengths of the FMF model is that it remains relatively consistent across multicenter external validations. In the prospective study of Cuenca-Gómez et al. (UOG 2024) including 10,110 pregnant women in Spain, the FMF model provided the best calibration for preterm preeclampsia with AUC of 0.911 (95% CI 0.879-0.943); in the same cohort, the NICE checklist method was able to detect only 46.7% of preterm PE at a screening-positivity rate of 11%, while the ACOG checklist was able to detect 65.9% at a high positivity rate of 33.8%. That is, the algorithmic competing-risks model works both more sensitively and more efficiently (with lower FPR) than traditional “risk factor counting” approaches.

ASPRE connection: where prediction meets prevention

The clinical value of a predictive model only emerges when it can trigger an accurate intervention. The evidence that solidifies the clinical rationale for FMF screening is the ASPRE study (Rolnik et al., NEJM 2017). In this multicenter, double-blind, placebo-controlled randomized study, 1776 women with singleton pregnancies identified as high risk by the FMF first-trimester model received aspirin 150 mg daily or placebo from 11–14 weeks until 36 weeks. The result was striking: preterm preeclampsia occurred in 1.6% in the aspirin arm and 4.3% in the placebo arm ( OR 0.38; 95% CI 0.20-0.74; P=0.004 ) — a reduction in relative risk of approximately 62%.

The paradigm brought by ASPRE is: algorithmic screening → selection of high-risk group → early aspirin prophylaxis chain can significantly prevent preterm preeclampsia. This is the strongest justification available for any predictive model: the model not only "predicts", it changes the outcome.

Machine learning-based models (2023-2026)

In recent years, machine learning models that learn directly from raw data have come to the fore as an alternative or complement to the competing-risks model. This approach has two main lines.

1. Biomarker-based neural network models

The first line is neural network models that use the same markers that FMF uses (MAP, UtA-PI, PlGF, PAPP-A), but learn from raw data without the need for MoM transformation. The fully-connected neural network tested in Gil et al.'s PREVAL cohort (Spain, 10,110 pregnant women; UOG 2024) detected 77.8% of preterm preeclampsia (AUC 0.913) and 84.4% of early (<34 weeks) preeclampsia at a 10% screening-positivity rate. This performance was broadly similar to the FMF competing-risks model in the same cohort (72.7% for preterm PE).

The finding that must be honestly emphasized here is that the ML model did not show a striking superiority over the FMF model; the two approaches were roughly equivalent. Moreover, in order to apply the model to a different population, a correction for the difference in the device (analyzer) measuring PlGF was necessary; without this correction, performance dropped significantly. This is a concrete example of how ML models cannot be brought to every center “out of the box”.

2. EHR/routine-data-based models

The second line, in systems where uterine artery Doppler is not routine (e.g., the USA), are models that learn from electronic health records (EHR) and routine pregnancy information. Kawakita and colleagues developed an XGBoost (gradient boosting tree algorithm)-based model with 131 variables available before week 23 in the nuMoM2b cohort (9,467 nulliparous pregnancies; Am J Perinatol 2025). The final model, reduced to eight variables, achieved an AUC of 0.779 (95% CI 0.722-0.831) in the validation set. This is significantly lower than the AUC of FMF-type biomarker models (≈0.91)—but in a context where Doppler and specific biomarkers are not routine, it is a reasonable performance achieved with currently collected data. The prominence of early second trimester uterine artery PI, chronic hypertension, and pregestational diabetes among the most predictive variables also indicates that the model is compatible with clinical logic.

Systematic reviews also confirm the general picture. Ranjbar et al. (4 studies; BMC Pregnancy Childbirth 2024) found the AUC range in ML models to be 0.860-0.973; Malik et al. (11 studies; Cureus 2024) reported it as 0.76-0.97. Both reviews pointed to XGBoost, random forest, and neural networks as prominent algorithms — but also clearly emphasized that a significant portion of the studies carried a high risk of bias.

Level of evidence and limitations

While high AUC values may be impressive at first glance, it is imperative to interpret performance figures honestly in this area. The most revealing data come from Liu et al.'s systematic review and meta-analysis (26 studies, 31 ML models; J Med Internet Res 2026):

• Although the pooled AUC was high at 0.91 (95% CI 0.87-0.92), the between-study heterogeneity was extreme (I²>99%)—that is, the models performed very differently in very different conditions.
• Only 6 studies had external validation; pooled sensitivity in these studies dropped to 0.68.
• The 95% prediction interval for sensitivity was very wide at 0.32-0.96 — meaning that in some outdoor environments the model's sensitivity could be as low as 32%.

The common message of these findings is clear: high AUC often reflects the “performance” of the model on the internal data set for which it was developed, and not its true “effectiveness” in independent and diverse populations. The apparent success of the model is highly context-dependent. When making clinical decisions, you should not look at a figure such as "0.97% AUC in the development cohort" of a model, but rather whether it has been externally validated in your own population.

If we list the limitations systematically:

• Transferability: A model trained on one population loses performance in another population with different ethnic distribution, biomarker measurement device, preeclampsia diagnostic criteria, and clinical workflow. Recalibration is often mandatory.
• Dataset imbalance: Preterm preeclampsia is a rare outcome (frequency 0.3-0.5% in cohorts); This imbalance makes classifier training and real-world positive predictive value difficult.
• Bias and inequality: Models may perform less well for groups underrepresented in training data (ethnic minorities, low-resource settings) — risking reinforcing existing inequalities in maternal health.
• Explainability: Some “black box” ML models pose problems in terms of clinical justification and ethical approval; therefore interpretability methods such as SHAP are increasingly in demand.
• Lack of prospective evidence: The vast majority of existing ML studies are retrospective; Prospective clinical impact studies, such as ASPRE, showing that it directly improves outcome (live birth, maternal morbidity) are still very limited.

Its reflection on clinical practice

Practical implications for the clinician from today's evidence base are:

• First trimester combined screening is a clinical reality, not a novelty. The FMF competing-risks model predicts preterm preeclampsia with higher sensitivity and lower false positivity than traditional risk-factor checklists (NICE/ACOG), and this prediction creates a concrete prevention opportunity by directing early aspirin prophylaxis, as demonstrated with ASPRE.
• For now, machine learning does not offer "revolution" but "equivalence and accessibility". Biomarker-based ML models perform largely equivalent to FMF; Their main contribution is that they do not require MoM conversion and — in EHR-based models — they can also be applied in systems where Doppler/specific biomarkers are not available.
• Local external validation is essential before a model is adopted. Published AUC values ​​cannot be directly extrapolated to your own population; Calibration may be required for device/population differences.
• Model output is a tool, not a decision. The decision for aspirin prophylaxis should be made within the context of clinical context, patient preference, and current guidelines (including dose/timing), along with the calculated risk score.

In summary, preeclampsia prediction is one of the most mature and well-demonstrated application areas of AI in obstetrics — but what drives this maturity is not the shiny AUC numbers, but the existence of a chain that has undergone multicenter external validation, such as the FMF model, and whose outcome-changing effect has been proven in a randomized trial such as ASPRE. New ML models must maintain this bar: they must be validated in independent populations, calibrated, and ultimately shown prospectively to improve outcome.

← Return to Artificial Intelligence in Obstetrics