Medical Technology · global
Looking for Preeclampsia Warning Signs in Retinal Blood Vessels: Retina-PE Completes Multicenter Data Validation
Combining retinal photographs, clinical data from pregnant women, and blood pressure, Retina-PE aims to identify the risk of subsequently developing preeclampsia. A study of 8,616 women provides evidence of predictive performance, but whether it can improve obstetric care still requires validation in clinical practice.
Preeclampsia can endanger pregnant women and fetuses during the middle and later stages of pregnancy; identifying risk early could provide more time for closer monitoring. A study published on November 17 in *npj Digital Medicine* turned its attention to the retina: the research team developed Retina-PE to test whether retinal photographs combined with clinical information could help predict the likelihood of subsequently developing preeclampsia.
The concept has a basis in vascular biology. Preeclampsia involves systemic vascular dysfunction, while the retina allows researchers to observe small blood vessels noninvasively. Imaging features such as vessel caliber and tortuosity may reflect changes in maternal circulation. However, Retina-PE does not rely solely on a single retinal photograph; it also incorporates maternal clinical factors and mean arterial pressure, which reflects blood pressure status.
The study used multicenter data from 8,616 pregnant women. A conference abstract for the same study released by the Korean Ophthalmological Society states that participants were recruited at two hospitals in China from 2020 to 2024, and that the model integrated imaging and clinical features through a cross-attention mechanism. A key distinction in understanding the study design is that, although the data were collected prospectively, model development and evaluation were conducted retrospectively. This was not a trial that deployed AI in outpatient care and then tracked physicians’ decisions and pregnancy outcomes.
According to the journal paper, in independent external validation data, the image-level AUCs for assessing subsequent preeclampsia at 24–28 and 28–32 weeks of gestation were 0.728 and 0.815, respectively; the AUC for predicting preterm preeclampsia within four weeks was 0.801. AUC measures a model’s ability to distinguish between those who develop the condition and those who do not. A value of 0.5 is close to random discrimination, while 1 represents perfect discrimination; 0.815 does not mean that 81.5% of pregnant women received a correct diagnosis.
These figures also cannot be treated directly as the results of screening each of the 8,616 pregnant women individually. The main results came from image-level analyses. The paper states that analyses at the individual level showed similar trends, but some prediction tasks involved very few disease events, limiting the certainty of the estimates. How many cases would actually be missed and how many false positives would occur also depends on the decision threshold used.
The added value of retinal imaging likewise needs to be considered separately for each task. Both the paper and reporting by Tıbbiye Bülteni indicate that the extent of improvement from adding images varied by prediction task; it cannot be inferred that all settings would benefit equally. Some AUCs listed in the conference abstract also differ from those in the journal version, so the figures used here are from the journal paper, without mixing the two sets of results.
Retinal photography therefore offers an approach to maternal risk assessment that merits further validation, but a gap remains between predictive performance and benefits to care. Prospective studies of clinical use are needed next to establish whether the model is reliable across different populations and healthcare settings, and whether arranging care based on its results can improve maternal and infant outcomes. Current evidence is insufficient to support replacing existing prenatal examinations with the model, or changing follow-up and treatment solely on the basis of its results.