20/20 Foresight

A handheld retinal camera with built-in AI detected diabetic retinopathy with 93.6% sensitivity using a single image per eye: no dilation, no specialist in the room.

The study, published in the Journal of Retina and Vitreous, tested the Eyer camera (Phelcom Technologies) on 686 people with diabetes in Blumenau, Brazil. Many were getting their retinas examined for the first time. A healthcare worker points the device, it captures a 45° image centered on the macula, and the onboard AI, EyerMaps, returns a result.

One image instead of the standard two. That’s the design choice that matters most here. Traditional screening protocols ask for multiple photos to cover enough of the retina. Eyer bets that a single, well-centered shot is sufficient for a screening-level decision. For mass screening in clinics without an ophthalmologist, that tradeoff (speed and simplicity over completeness) makes sense.

The specificity landed at 71.7%, which means roughly three in ten people without disease would get flagged for follow-up they don’t need. In a screening context, that’s acceptable. You’d rather over-refer than miss someone. But it’s worth watching as the system moves toward broader deployment.

A couple of caveats: all images were graded against a single retinal specialist, not a panel, and ungradable images were excluded from the analysis. Both of those choices flatter the numbers. The device still lacks FDA clearance as an autonomous diagnostic, so for now it operates as a screening aid.

Still, a first retinal exam for hundreds of people who’d never had one. That’s not a future promise. That already happened.

Seven years. That’s how far in advance 3D retinal scans picked up changes associated with Parkinson’s disease, according to a study from University College London and Moorfields Eye Hospital published in Nature Medicine.

The team analyzed OCT scans - the same quick, non-invasive imaging you get at a routine eye appointment - from over 220,000 people, pulling data from the AlzEye dataset and the UK Biobank.

What they found: people who went on to develop Parkinson’s already had measurably thinner ganglion cell and inner plexiform layers, plus specific thinning in the inner nuclear layer.

The retina, it turns out, is a remarkably readable window into neurodegeneration. The same cell loss happening in the brain shows up in the back of the eye years earlier, in a scan that takes seconds.

This isn’t a diagnostic yet. It’s a signal. But it’s a signal detectable with equipment that already sits in thousands of optometry offices worldwide. The research team is now working to validate a foundation model that could scale this across clinical settings, part of a growing field called oculomics, using eye data to surface conditions that have nothing to do with vision.

The practical upside: if you can flag risk at the pre-symptomatic stage, you can start interventions (exercise, medication, lifestyle changes) when they’re most likely to matter. Parkinson’s is typically diagnosed after motor symptoms are already obvious.

A seven-year head start changes the math on what’s possible.

Eighteen blind and low-vision users tried navigating with standard GPS. They succeeded 38% of the time. Then they used NaviNote, an AI-powered tool that lets users create voice-tagged spatial annotations as they walk. Success rate: 88%.

NaviNote solves a specific problem that GPS doesn’t: the last few meters. Standard navigation gets you to the block. NaviNote, using high-accuracy positioning, gets you to the door and remembers what you told it about that door the last time you were there. Every annotation is user-generated, which means the system gets better the more someone uses it in their own environment.

It’s not the only tool making progress.

- Mobilio, an audio-guided navigation app, cut navigation time by 13% and reduced unintended collisions with the environment by 41% in outdoor tests.

- SnapStick, a smart cane with AI object recognition, hit 94% accuracy identifying obstacles and scored 84.7% on usability, which UX researchers consider exceptional.

- LLM-Glasses, which use haptic feedback (vibrations, not audio) to guide directional decisions, achieved 91.8% accuracy in open environments.

What connects all four: they’re designed around how blind and low-vision users actually move through space, not how sighted engineers imagine they do. NaviNote lets users build their own navigation knowledge. SnapStick works through a form factor people already carry. Mobilio prioritizes collision avoidance, not just routing. These aren’t generic GPS wrappers with accessibility bolted on.

In 2018, the FDA cleared IDx-DR as the first fully autonomous AI diagnostic in any field of medicine. No physician reads the image. The software makes the call. The manufacturer carries the liability.

The pivotal trial enrolled 900 adults across 10 sites. The bar: 85% sensitivity, 82.5% specificity, measured against widefield stereoscopic photography read by masked graders. IDx-DR hit 87.4% and 89.5%. Two more systems followed the same pathway - EyeNuk’s EyeArt, and AEYE Health’s AEYE-DS, which added a handheld camera and single-image protocol.

All three do the same thing: a medical assistant in a primary care clinic takes a retinal photo. The AI returns a binary result in under a minute: referable diabetic retinopathy detected, or not. If image quality fails, it says so.

The regulatory infrastructure the FDA built around this is worth understanding. A new product code (QFM) specifically for autonomous AI diagnostic software. A reimbursement pathway via CPT code 92229, paying roughly $50 per screen. The designation means the device makes the clinical decision, not the clinician - a distinction that didn’t previously exist in FDA product classification.

Here’s the number that puts all of this in context: clinical guidelines say every diabetic patient should get an annual retinal screen. Historically, 40-50% actually do. The gap isn’t a technology problem or an awareness problem. It’s a logistics problem - getting a patient in front of the right specialist with the right equipment. These devices sit in family medicine clinics. They don’t require an ophthalmologist. AEYE-DS doesn’t even require dilation.

Three devices, one regulatory pathway, and a template for how autonomous clinical AI gets to market.