← Back to Home

Predicting Preeclampsia From Retinal Blood Vessels: Retinal AI Passes Independent External Validation

Pregnancy risks may leave microvascular clues before symptoms appear. Visionary AI retained predictive ability in early-pregnancy images from another hospital, but this small validation study still leaves clinical benefits to be demonstrated before routine screening.

By SURL BioNews

Preeclampsia often shows clinical signs only after 20 weeks of pregnancy, but changes in vascular function may begin earlier. Identifying risk in advance could provide more time to arrange follow-up in prenatal care. A study published on October 6 in *Nature Biotechnology* turned to the retina for clues: Visionary AI used retinal vascular features to predict the risk of subsequently developing preeclampsia and completed independent external validation.

The retina provides a window into the microvasculature. The system first identifies blood vessels in images, then quantifies structures such as branching, connectivity, and curvature to generate predictions. This allows researchers to trace which vascular features are associated with risk; however, the interpretability of these features does not establish that they cause the disease.

The study's prospective development cohort included 1,267 pregnant women. In analyses using sampled controls and cross-validation, the model optimized for performance within the hospital achieved an AUC of 0.91 for predicting preeclampsia. AUC measures a model's ability to distinguish between people who subsequently develop the disease and those who do not. It cannot be read directly as “91% accuracy,” nor can this number alone indicate how many false positives would occur in actual screening.

A step closer to testing practical application was transferring the model focused on stability to an independent NYU cohort, without retraining or using local outcomes to adjust the model, yielding an AUC of 0.81. This analysis used images from 6 to 13 weeks of pregnancy and tested the model's performance across different institutions, populations, and imaging equipment. However, the cohort included only 79 people, with just 13 preeclampsia cases actually included in the image analysis.

NYU Langone's research overview provided further details on the clinical setting for data collection: the team arranged retinal photography and blood sampling during routine prenatal visits in all three trimesters, then tracked pregnancy outcomes through medical records, and listed Liat Shenhav and Justin Brandt as the principal investigators. This describes the research process and does not justify assuming that the retinal model's performance came from blood testing. The Shenhav laboratory also confirmed that the paper was published on October 6 but provided no additional performance or patient outcome data.

External validation makes the results more persuasive, but the sample size still limits the conclusions that can be drawn. The proportion of preeclampsia cases was higher in the NYU cohort, and disease prevalence in a screening population affects how many positive results correspond to people who actually develop the disease. The study also covered only two US academic medical centers; performance still needs to be verified in other regions, different risk groups, or with lower-cost equipment.

The next key question is how to incorporate risk scores into care: at what threshold to arrange additional assessment, how to handle false positives, and whether this can improve maternal and infant outcomes. This study supports further evaluation of retinal screening in early pregnancy but has not demonstrated that acting on AI results can reduce complications. Larger prospective studies are still needed to determine whether retinal imaging can become part of prenatal care.

References

  1. Nature Biotechnology, Published online: 2026-10-06; | doi:10.1038/s41587-026-03303-0
  2. Shenhav Lab
  3. NYU Langone Health