The Eye as a Biomarker: Why Retinal AI Could Reshape How Trials Screen and Stratify Pregnant Patients

Picture a study coordinator at a maternal-fetal medicine clinic, running a retinal scan on a patient at 12 weeks gestation. No blood draw. No invasive uterine artery Doppler. No specialist referral. The image uploads, an algorithm processes the microvascular geometry captured in retinal imaging, and a risk score appears in the eClinical dashboard. High probability of preeclampsia. Flag for enhanced monitoring. The trial’s adaptive stratification logic triggers automatically, routing the participant into a more intensive intervention arm before a single symptom has appeared.

That scenario moved meaningfully closer to clinical reality with the publication in Nature Biotechnology of Visionary AI, a deep-learning framework trained to decode systemic vascular health and hypertensive disorders in pregnancy through retinal imaging. The clinical stakes alone justify attention: gestational hypertension affected 10.4% of U.S. births in 2024, a 73% increase from the 6.0% rate recorded in 2016, with absolute case counts rising from 235,693 to 377,267 in that eight-year window. But the implications for clinical trial design extend well beyond obstetrics, into the architecture of how sponsors stratify, monitor, and generate real-world evidence across virtually any cardiovascular or metabolic indication.

A Window Into Vascular Biology

The retina is, anatomically, an extension of the brain, the only place in the human body where microvasculature can be observed non-invasively, in real time, without a catheter or contrast agent. That fact has long attracted researchers trying to use fundus photographs as a proxy for systemic vascular disease. What Visionary AI adds is scale and precision. Earlier validation work published in PMC drew on 95,665 macula-centered retinal images from 51,778 UK Biobank participants to train deep-learning models that predicted systolic blood pressure directly from fundus images, with no prior cardiovascular disease in the cohort. The Visionary AI paper extends that logic into pregnancy, where vascular dysregulation arrives faster, matters more, and currently gets detected far too late.

The regulatory environment for models like this is evolving in parallel, but not at the same speed as the science. The FDA’s Center for Devices and Radiological Health finalized guidance in December 2024 on Predetermined Change Control Plans for AI-Enabled Device Software Functions, which allows manufacturers to prospectively define how an AI model can be updated without triggering a full de novo review each time. That guidance matters enormously for a platform like Visionary AI, where the model will almost certainly need recalibration as it encounters population-level demographic variation, equipment heterogeneity across sites, and longitudinal drift in the training distribution.

But the PCCP guidance was written for cleared devices used in routine care. Visionary AI, applied inside a clinical trial as a patient stratification or enrollment eligibility tool, sits in a regulatory gray zone that sponsors have not yet fully mapped. That gap deserves a closer look.

The Stratification Opportunity, and the Regulatory Gap

Here is the concrete operational scenario that every sponsor running a pregnancy or cardiovascular trial should be thinking through right now. A retinal AI tool used at screening to predict hypertensive disorder risk would function as a biomarker-based enrichment strategy: enroll patients above a defined risk threshold, exclude those below it, reduce placebo arm event rates to something that actually powers the study. The FDA has long encouraged biomarker-driven enrichment; the 2019 Enrichment Strategies for Clinical Trials guidance makes that explicit. The problem is validation evidence.

The preeclampsia prediction literature shows why validation rigor matters. A 2022 retrospective case-control study published in the Annals of Translational Medicine developed a Random Forest model for preeclampsia prediction with 79.6% sensitivity and 94.7% specificity. Impressive on paper. But a 79.6% sensitivity in an enrichment context means roughly one in five high-risk participants gets routed to the control arm incorrectly, a contamination problem that inflates variance and can hollow out a trial’s statistical power before the first dose is administered. Sponsors considering retinal AI for stratification need prospective validation data, not retrospective case-control performance metrics, and they need it in a population that matches their trial cohort.

The CDRH’s updated December 2025 guidance on Real-World Evidence for Medical Devices creates a pathway, but it also sets the bar. The guidance requires that RWE used to support regulatory decisions meet standards for data provenance, pre-registration, and transparent methodology. Retinal imaging data collected opportunistically at antenatal clinics, without a pre-specified analysis plan and without documentation meeting applicable data integrity requirements, will not qualify. The window of the eye is only as useful as the governance around what gets captured through it.

Which raises the harder question: what would a properly designed Visionary AI validation study actually require before a sponsor could use this tool to gate trial enrollment?

What Sponsors Must Demand Before Deploying This at Scale

Consider what Cleveland Clinic demonstrated in a parallel domain. An AI system deployed within their health system for rare disease trial eligibility screening achieved 96% accuracy across nine prespecified eligibility domains, identifying 46 potential matches of whom 43 were confirmed appropriate after human review. That result sounds compelling, but the key phrase is “within a health system firewall.” The model performed against a population it had, in effect, already seen. Retinal AI for pregnancy stratification will face a more demanding challenge: it must generalize across imaging equipment from different manufacturers, across maternal populations with different baseline prevalence rates for hypertensive disorders, and across trial sites where fundus photography quality varies considerably.

The AI ophthalmic diagnostics market was valued at USD 286.8 million in 2025 and is projected to grow at 36.7% CAGR through 2033. That growth reflects genuine clinical demand. But it also reflects a market moving faster than the evidentiary standards required to deploy these tools inside regulated trials. A sponsor who integrates a retinal AI stratification tool without pre-specifying the algorithm version, locking it for the duration of the trial, and validating its performance in a bridging study on the actual trial population is building enrollment architecture on an unvalidated foundation.

Three things a sponsor must negotiate before the first site activation. First, a locked algorithm version with a defined performance specification tied to the imaging hardware used at each trial site, not a platform-level specification, but a site-level one. Second, a pre-specified algorithmic failure mode protocol: what happens when image quality falls below the model’s confidence threshold, and who makes the eligibility decision in that case? Third, an IRB-reviewed consent process that discloses to participants that a machine learning tool is influencing their trial assignment, a disclosure requirement that many current eConsent platforms may not be configured to surface clearly.

None of those requirements are unreasonable. All three are achievable. But they require sponsors to treat a retinal AI stratification tool the way they treat any other biomarker assay: as a regulated analytical procedure with a qualification package, not as a software feature that gets activated at go-live.

The coordinator with the fundus camera at that 12-week visit has a genuinely powerful instrument in her hands. Whether the trial she is supporting can use what it generates depends entirely on decisions made in the protocol, the statistical analysis plan, and the regulatory strategy months before she scans the first patient. Visionary AI has delivered the science. The clinical operations community now has to build the infrastructure to make it count.

References

1. Nature Biotechnology, “Decoding systemic vascular health and hypertensive disorders in pregnancy through retinal imaging and Visionary AI”
2. OPQIC / CDC, “Trends and Characteristics in Gestational Hypertension, United States, 2016–2024”
3. PMC, Retinal imaging deep-learning model for systolic blood pressure prediction, UK Biobank (95,665 images, 51,778 individuals)
4. Ropes & Gray, “FDA Finalizes Guidance on Predetermined Change Control Plans for AI-Enabled Device Software Functions” (December 2024)
5. FDA CDRH, “Real-World Evidence: Advancing Regulatory Decision-Making for Medical Devices” (updated December 2025)
6. Annals of Translational Medicine, “Development and validation of prediction models for preeclampsia” (Random Forest model; sensitivity 79.6%, specificity 94.7%; November 2022)
7. Cleveland Clinic Consult QD, “AI Can Unlock EHR Data to Determine Trial Eligibility” (96% accuracy across nine eligibility domains)
8. Metastat Insight, “Global AI Diagnostics in Ophthalmology Market” (USD 286.8 million, 2025; 36.7% CAGR through 2033; April 2026)