A cloud-based AI system that analyzes a snapshot of your retina without dilating your pupils can detect signs of potentially vision-threatening macular degeneration with over 90% sensitivity, according to a prospective clinical study published July 29 in Scientific Reports. The system, called iPredict-AMD, delivered results in roughly one minute and was tested in real-world primary care and ophthalmology clinics — a setting far closer to where most patients first seek care than the tightly controlled lab conditions of earlier research.
This isn’t a theoretical lab demo. The study enrolled 696 adults over 50 with no prior AMD diagnosis across six sites in the New York City area, including three primary-care clinics. Participants had their eyes imaged with a non-mydriatic fundus camera (a DRSPlus, which doesn’t require dilation), and the AI’s assessments were compared against in-depth evaluations by three independent ophthalmologists using dilated exams. The result: 90.27% sensitivity and 83.36% specificity for spotting “referable” AMD — that’s intermediate or late-stage disease that warrants a specialist’s attention.
The numbers behind the claim
At first glance, those figures sound impressive, and they are. But to understand what they mean for you, a few details matter. Sensitivity tells us how good the system is at catching people who truly have the condition. Out of 113 participants confirmed to have referable AMD by the ophthalmologist panel, the AI correctly flagged 102. The other 11 received a “negative” result and were advised to come back in a year for routine rescreening.
Those 11 misses are a reminder that no screening tool is perfect. Yet the study’s most reassuring number isn’t the sensitivity — it’s the negative predictive value (NPV). At 97.79%, when the AI said you’re fine, it was almost always correct. In other words, a “normal” report is highly trustworthy. That matters in a primary-care setting where false alarms can create anxiety and unnecessary referrals.
There’s a twist, though. The AI was designed to look only for AMD, but it flagged several participants with other retinal problems as “referable” anyway. Among 24 people with referable diabetic retinopathy, 15 were caught; all participants with epiretinal membrane, high myopia, or macular pucker were flagged. The researchers call this “an added benefit,” but they’re careful to note the system wasn’t validated for those conditions. If your scan comes back positive, it means “see an eye doctor,” not a specific diagnosis.
Who stands to gain — and what’s still missing
For anyone over 50 who dreads the sting of dilation drops and the blurry drive home, the promise is obvious: a quick, painless photo that could be taken during a routine checkup and analyzed while you wait. Age-related macular degeneration is a leading cause of vision loss, and catching it early can make a difference. The landmark AREDS study showed that certain vitamins and lifestyle changes can reduce the risk of progression from intermediate to late AMD in about 25% of patients. Yet many people don’t see an ophthalmologist until symptoms appear. A tool embedded in a primary-care visit could close that gap.
But it’s not yet ready for your next annual physical. The study, though prospective and multi-site, was limited to New York City. Of 845 original enrollees, 149 dropped out or didn’t complete the protocol. The underlying dataset belongs to iHealthscreen, a company affiliated with some of the study’s authors. And while the system has been submitted for FDA 510(k) clearance — as announced in May 2023 — it isn’t approved yet. A larger pivotal trial is ongoing with an estimated completion date of July 2027, so regulatory green light and broad availability are still years away.
For primary-care physicians, the tool represents a new workflow challenge. The camera itself (the DRSPlus) is relatively low-cost and user-friendly enough for office staff to operate, but the AI analysis happens in the cloud. That means a clinic needs reliable broadband, secure image transfer, and a way to integrate results into the electronic health record. The study’s authors explicitly mention barriers in rural and low-resource areas: image quality can degrade, internet can be spotty, and staff need training. None of that is insurmountable, but it’s not plug-and-play.
Health IT administrators in clinics that run on Windows endpoints will face familiar tasks: ensuring browsers meet security requirements for cloud-based medical devices, managing user authentication, maintaining HIPAA compliance for data in transit and at rest, and planning for downtime or connectivity outages. iPredict-AMD’s cloud architecture means the heavy lifting happens offsite, but that also means the local workstation becomes a critical link. If the upload fails or the report doesn’t return promptly, the screening cascade stalls. Early adopters will need to test these workflows thoroughly before relying on the system for patient care.
The long path to a one-minute screen
iPredict-AMD didn’t appear overnight. Its underpinnings are an ensemble of five deep-learning networks built on well-known architectures — Xception, Inception-V3, and Inception-ResNet-V2 — trained to spot the subtle retinal signs of AMD. Before this prospective study, the team validated the model against a retrospective dataset, and results were published earlier. A separate FDA-cleared device, AEye-ds, paved the way by demonstrating that AI retinal screening could pass regulatory muster, though it was tested on fewer than 500 subjects.
The idea of automated retinal screening isn’t new. Google, IDx-DR, and others have developed AI for diabetic retinopathy, some with FDA clearance. AMD screening has historically been trickier because the disease markers are more varied. What sets iPredict-AMD apart is its focus on non-dilated imaging and its explicit design for primary care — a setting where most previous tools weren’t tested.
The business push is real. In 2023, iHealthScreen submitted its 510(k) application, framing the tool as software-as-a-medical-device. That classification means it’s subject to regulatory oversight, but also that updates can be iterated more easily than with dedicated hardware. The ongoing pivotal trial is designed to support that clearance and, presumably, to satisfy insurers that the screening is cost-effective. The paper argues that the modest expense of screening would be offset by future savings from preventing advanced AMD, but no formal cost analysis is included.
What you can do right now
If you’re over 50 and worried about macular degeneration, the best step today remains a comprehensive eye exam with dilation — especially if you have risk factors like smoking, family history, or light-colored eyes. Ask your primary-care doctor if they offer retinal photography and what they do with the images. Some practices already use fundus cameras to document diabetic eye disease; the leap to AI-assisted screening isn’t large, but it requires a validated tool and a workflow commitment.
For clinic managers and IT decision-makers, this study is a signal to start evaluating your infrastructure. Check your broadband reliability. Review how your EHR system handles external reports. Talk to your staff about their comfort with ophthalmic imaging. Even if you’re not in the market for iPredict-AMD today, similar tools are coming. The FDA’s recent clearances in retinal AI suggest that primary care will increasingly be the frontline for eye disease detection.
If you work in a rural or underserved clinic, pay attention to the upcoming trial results and to any companion efforts to make the technology offline-capable or operable with portable cameras. The researchers acknowledged these gaps, and addressing them will be essential for equitable deployment. In the meantime, teleophthalmology programs that use human graders can still fill the gap.
What comes next
The 2027 trial completion date is the nearest lighthouse. If results are positive and the FDA clears the device, we can expect a gradual rollout, likely starting with larger health systems that have the IT muscle to integrate it. Version updates may expand to explicitly cover diabetic retinopathy or glaucoma — the fact that the AMD tool snagged some of those cases suggests the underlying models can be tuned.
The authors call out EHR integration as a critical next step. That will require standards-based interoperability, likely using FHIR APIs, to automatically push results into patient records and trigger referral orders. For Windows-centric clinics, that might mean integration with common systems like Epic or Cerner via browser plugins or dedicated agents.
For now, the takeaway is clear: AI is moving from bench to bedside in eye care. The one-minute, no-drop retinal scan isn’t science fiction. It’s been tested on real patients in real clinics, and it worked well enough to change how we think about AMD screening. The wait is on for the regulatory rigor that will decide when — and where — it becomes part of your annual checkup.