Retinal AI Predicts Preeclampsia Before Symptoms in Small External Study

Images of retinal blood vessels can help predict preeclampsia before symptoms appear, researchers report in an October 6 Nature Biotechnology paper. Their Visionary AI framework carried that predictive signal into an independent hospital cohort without retraining. The result points toward earlier pregnancy-risk assessment, although the external retinal analysis included only 13 preeclampsia cases.

Reading the vessel network, not just the image

Visionary AI turns retinal images into measurements of blood-vessel shape and connections. Its features describe the network’s geometry, complexity and hierarchical organization. The researchers designed it to make predictions from identifiable vascular properties, rather than generic image representations or clinical variables.

That design follows the study’s biological premise: changes in the retina’s small vessels may reflect vascular stress elsewhere in the body during pregnancy. Hypertensive disorders are typically diagnosed after 20 weeks, once high blood pressure or other symptoms become apparent. The researchers aim to detect earlier changes that precede those clinical signs by weeks or months.

A transfer between hospitals

The prospective development cohort covered 1,267 pregnancies at Columbia University between 2021 and 2025. Fifty-five participants developed preeclampsia; one was excluded from retinal analysis because of poor image quality. The paper reports an area under the curve, or AUC, of 0.91 and average precision of 0.81 for preeclampsia in the development cohort.

For the independent NYU test, the team used a simplified model optimized for stability. It was neither retrained nor optimized using NYU outcome labels. That cohort included 79 pregnancies from 2025–2026, with 14 preeclampsia diagnoses before one image-quality exclusion. The researchers report that it outperformed the study’s clinical and deep-learning benchmarks.

0.81Preeclampsia AUC

The stability-optimized model achieved AUC 0.81 in the independent NYU validation cohort.

0.68Average precision

The same external evaluation returned average precision of 0.68.

Both hospitals collected images across all three trimesters using Optos ultrawidefield imaging systems. Preeclampsia diagnoses were independently reviewed by maternal–fetal medicine specialists who were masked to the retinal results. The NYU cohort’s observed preeclampsia prevalence was 17.7%, above the hospital’s roughly 8% annual average. The authors note that prevalence affects precision-based metrics and report prevalence-adjusted estimates.

Trial screening remains a proposed use

In October 8 commentary for Clinical Trial Vanguard, Moe Alsumidaie proposed using retinal AI to screen trial participants and group them by risk. That is a possible application, not an established Visionary AI trial workflow. He argues that sponsors would need prospective validation in a population matching their trial before using the tool to influence enrollment.

Alsumidaie also calls for a fixed algorithm version throughout a trial and a defined procedure for images that fail quality requirements. Those recommendations address how a risk score would become a consequential decision: which participants enter a study, how they are assigned, and who decides when the image cannot support the automated assessment.