The stakes are considerable. Hypertensive disorders of pregnancy, which include gestational hypertension, preeclampsia, chronic hypertension with superimposed preeclampsia and eclampsia, affect between 2 and 15 percent of pregnancies and rank as the second leading cause of maternal mortality worldwide. They sharply raise the risk of acute maternal complications such as renal dysfunction, pulmonary edema, stroke and death, and they contribute to adverse fetal outcomes including small-for-gestational-age infants, preterm delivery and stillbirth. The consequences reach far beyond delivery: women who experience these disorders carry elevated lifetime risks of hypertension, coronary artery disease, stroke and vascular dementia, and their children face higher risks of cardiovascular and neurodevelopmental problems. In the United States alone, hypertensive disorders of pregnancy were estimated to cost the healthcare system 2.18 billion dollars per year as of 2012, largely driven by complications of preterm birth.

What makes the new work compelling is its biological rationale. Preeclampsia is believed to arise in part from maternal vascular inflammation and thrombosis that disrupt placental development, impairing the deep invasion of placental tissue into the maternal decidua between roughly 10 and 16 weeks of gestation. These early deviations from normal placentation precede clinical symptoms by weeks or months, and they share features with the vascular pathology seen in cardiovascular disease. Because the retina is the only place in the body where capillaries, arterioles and venules can be directly and noninvasively visualized, the researchers reasoned that the maternal eye could serve as a living biosensor of the systemic vascular stress that accompanies early placental dysfunction. Prior studies had already found increased retinal arterial tortuosity in women with severe preeclampsia, narrower vessels and reduced capillary density in women who later developed placental insufficiency, and lower vessel densities in both the superficial and deep capillary plexuses of women with the disease.

Current screening, however, remains blunt. Preeclampsia is typically diagnosed only after 20 weeks of gestation, when hypertension, proteinuria or organ dysfunction become clinically apparent. Known risk factors such as prior hypertensive disorders, diabetes, obesity and advanced maternal age provide some predictive value but lack individual-level accuracy. The most validated biomarker, the ratio between the antiangiogenic factor sFlt-1 and placental growth factor, is primarily useful later in pregnancy and in women already suspected of having the disease. First-trimester tools such as the Fetal Medicine Foundation risk calculator combine clinical variables with biomarkers or ultrasound measurements, but these protocols are expensive, logistically complex and inconsistently implemented. In routine practice, prevention-oriented risk stratification often reduces to deciding who qualifies for low-dose aspirin, even though randomized trials show that starting aspirin before 16 weeks can reduce the risk of preterm delivery with preeclampsia by up to 62 percent. Accurate, early identification is the bottleneck that prevents that intervention from reaching the women who need it most.

Visionary AI attacks the problem with an architecture designed for interpretability rather than brute-force pattern recognition. The pipeline begins with ultrawidefield retinal images, captured by scanning laser ophthalmoscopes that cover roughly 80 percent of the retinal surface. A dedicated vessel segmentation model, inspired by generative adversarial network architectures and fine-tuned on hand-annotated ultrawidefield images, converts each photograph into a binary map of the vascular tree. When multiple images of the same eye exist, the segmentations are merged using keypoint matching and deformable registration to recover regions obscured by eyelashes or shadows. The resulting vessel maps are then transformed into three complementary families of features: graph-based measures of branching and connectivity, geometric measures of vessel length, caliber, curvature and tortuosity, and higher-order topological descriptors, including box counting, fractal organization and topological data analysis, that quantify loop structures and hierarchical network organization across scales.

These feature sets feed a stacked ensemble of machine learning models. Each semantically grouped feature set trains independent base learners built on logistic regression, random forests and extreme gradient boosting, and their out-of-fold predictions are integrated by a metalearner. The team evaluated the framework across increasingly stringent settings. In a high-contrast analysis comparing 54 preeclampsia cases against 82 carefully screened healthy controls, the model achieved an AUC of 0.94 and a positive predictive value of 0.85. When the comparison was broadened to a clinically heterogeneous population-wide control pool that deliberately retained women with obesity, diabetes and chronic hypertension, performance remained strong at an AUC of 0.91 plus or minus 0.05. Notably, adding the Fetal Medicine Foundation calculator’s probabilities to the retinal predictions did not materially improve performance, suggesting that the retinal vasculature captures most of the relevant risk signal on its own.

The most demanding test was external. The researchers built a stability-optimized model that prioritized reproducibility over peak performance, using stability-based hyperparameter selection, recursive feature elimination and feature-set deduplication to shrink an initial search space of 96 candidate base learners down to just eight nonredundant learners per fold. This deliberately constrained model was then applied, frozen and untouched, to an independent cohort of 79 pregnant individuals recruited at NYU Langone Health, a population that differed from the development cohort in race and ethnicity distribution, maternal age, in vitro fertilization prevalence and imaging hardware. Restricted to first-trimester images acquired between 6 and 13 weeks of gestation, the model achieved an AUC of 0.81 and an average precision of 0.68, outperforming the official Fetal Medicine Foundation calculator, which reached an AUC of 0.73 and an average precision of 0.50 on the same cohort.

The benchmark against routine clinical practice is perhaps the most striking result. At NYU, aspirin eligibility based on standard clinical criteria identified preeclampsia cases with a sensitivity of 0.62 but at a false positive rate of 0.47 and a prevalence-adjusted positive predictive value of just 0.1. Visionary AI achieved a higher sensitivity of 0.69 at a false positive rate of only 0.13, with a prevalence-adjusted positive predictive value of 0.31. Even within the subgroup of women already recommended aspirin, a clinically enriched high-risk group, the retinal model maintained an AUC of 0.95. The framework also outperformed two direct image-level deep-learning approaches, the retinal foundation model RETFound and a convolutional network, both of which performed near chance in external validation, a finding the authors interpret cautiously but which supports the idea that biologically informed, lower-dimensional vascular representations are better suited to clinical cohorts with limited numbers of outcome events.

Beyond prediction, the features themselves tell a coherent biological story. Recurrently selected features involved retinal vascular topology, geometry, network complexity and hierarchical organization, and model-independent analyses confirmed preeclampsia-associated differences in branch count, topological length, tortuosity, vessel thickness, fractal organization, loop count and vascular asymmetry. The signatures point to vascular simplification through two complementary processes: capillary rarefaction, the loss of small vessels and reduced branching complexity, and arteriolar wall thickening that narrows luminal diameter. Intriguingly, gestational and chronic hypertension produced distinct retinal fingerprints, with localized simplification of hierarchical vascular organization, whereas preeclampsia showed more globally extensive sparsity and disorganization. The authors stress that these associations are hypothesis generating rather than causal, and that the validation cohort, while genuinely independent, remains modest in size with a higher observed preeclampsia prevalence than the institutional average.

The practical appeal is hard to overstate. Retinal fundus imaging requires no laboratory infrastructure, takes less than two minutes, and can be performed by trained nonspecialists, making the approach potentially viable in both high-resource and low-resource settings. The computational pipeline is fully automated, eliminating subjective image grading. Before Visionary AI reaches the clinic, larger and more geographically diverse prospective studies will be needed to establish calibration, optimal thresholds and real-world screening performance, and ongoing work aims to integrate the retinal signal with circulating biomarkers and placental pathology. But if the findings hold, the humble eye photograph, already a staple of ophthalmology and increasingly a window into diabetes and Alzheimer’s disease, may become one of the most consequential tools in prenatal medicine: a two-minute scan that flags silent vascular danger early enough for aspirin, surveillance and time to change the outcome.

Subject of Research: Retinal imaging and interpretable AI for predicting hypertensive disorders of pregnancy

Article Title: Decoding systemic vascular health and hypertensive disorders in pregnancy through retinal imaging and Visionary AI

Article References: Bearelly, S., Areff, C., Hoffman, M. K., Sunko, D., Laine, A. F., Choi, B., Coleman, H. R., Cordazzo Vargas, B., Li, C. Y., Dawkins, J. J., Amir, E., Paz, S., Hark, L. A., Booker, W. A., Jauhal, A. A., Brandt, J. S., Wapner, R. J., & Shenhav, L. (2026). Decoding systemic vascular health and hypertensive disorders in pregnancy through retinal imaging and Visionary AI. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03303-0

Image Credits: AI Generated

DOI: 10.1038/s41587-026-03303-0

Keywords: preeclampsia, hypertensive disorders of pregnancy, retinal imaging, artificial intelligence, Visionary AI, vascular topology, predictive biomarkers, maternal health, machine learning, prenatal screening, placental dysfunction, Nature Biotechnology