LumineticsCore vs EyeArt vs AEYE-DS vs iPredict-DR: Autonomous AI Diabetic Eye Exams Compared (2026)

Most of the AI we cover on this site drafts something for a clinician to sign. This category is different. An autonomous diabetic retinopathy (DR) system takes a retinal photo in a primary care office and issues the result itself, with no ophthalmologist or optometrist reading the image. It has its own CPT code, it counts toward a HEDIS quality measure, and as of July 2026 there are four FDA-cleared products instead of three.

It is also barely used. This comparison covers what each system is actually cleared to do, which camera you are committing to, what the trial numbers mean in a real waiting room, and why adoption has stalled.

The short version

• Four systems hold FDA clearance under the same product code (PIB): LumineticsCore (formerly IDx-DR), EyeArt, AEYE-DS, and the newcomer iPredict-DR, cleared July 2, 2026.
• Each clearance is tied to specific cameras. You are choosing a camera-plus-software pair, not an app.
• Only AEYE-DS is cleared with a handheld camera. Only EyeArt is cleared to flag vision-threatening DR as a separate result.
• No study has compared the systems on the same patients. The sensitivity and specificity figures below come from separate trials and should not be ranked against each other.
• None of the vendors publish pricing. Medicare pays for the service under CPT 92229, in the $40s nationally.
• A positive result is a referral, not a diagnosis. In a typical primary care population, roughly half of positives will be false alarms.

Why this exists

The American Diabetes Association recommends an eye exam for people with type 2 diabetes at diagnosis and generally every year after that (every one to two years if prior exams were clean and glycemia is at goal). Many patients never make it to the eye doctor. The pitch for autonomous AI is simple: do the screen during the diabetes visit the patient is already at.

The ADA's 2026 Standards of Care back the approach. Recommendation 12.6 reads:

"Programs that use retinal photography with remote reading or the use of U.S. Food and Drug Administration–approved artificial intelligence algorithms to improve access to diabetic retinopathy screening are appropriate screening strategies."

The same section adds the caveat that matters most for your workflow: "In-person exams are still necessary when the retinal photographs are of unacceptable quality and for follow-up if abnormalities are detected. Retinal photographs are not a substitute for dilated comprehensive eye exams."

The four systems

All four share the core indication: use by healthcare providers to automatically detect more than mild diabetic retinopathy (mtmDR) in adults with diabetes who have not previously been diagnosed with DR. The differences are in cameras, outputs, and evidence.

LumineticsCore (Digital Diagnostics)

• Regulatory history: De Novo DEN180001, granted April 11, 2018, as IDx-DR. It was the first autonomous AI diagnostic FDA authorized in any field. Later 510(k)s: K203629 (June 2021) and K213037 (June 2022).
• Camera: Topcon NW400 only, a tabletop non-mydriatic camera.
• Pivotal trial: 900 patients in primary care offices (Abràmoff et al., npj Digital Medicine, 2018). Sensitivity 87.2%, specificity 90.7%, imageability 96.1%. Operators were existing staff given a one-time four-hour training. 23.6% of participants needed dilating drops to get a result.
• Worth knowing: The vendor's labeling is for adults 22 and older and lists who should not be screened: patients with visual symptoms, pregnant patients, and anyone with prior retinal laser, injections, or surgery. It has the deepest independent literature of the four.

EyeArt (Eyenuk)

• Regulatory history: K200667 (August 3, 2020), then K223357 (June 16, 2023) for version 2.2.0.
• Cameras: Canon CR-2 AF, Canon CR-2 Plus AF, and Topcon NW400. All tabletop.
• Pivotal trial: 893 patients at 15 primary care and eye care sites (Ipp et al., JAMA Network Open, 2021). For mtmDR: sensitivity 95.5%, specificity 85.0%. For vision-threatening DR: sensitivity 95.1%, specificity 89.0%. Both without dilation.
• Worth knowing: EyeArt is the only one of the four cleared to report vision-threatening DR as its own category, which helps you triage an urgent referral from a routine one. Its indication is written per eye.

AEYE-DS (AEYE Health)

• Regulatory history: K221183 (November 10, 2022) with the tabletop Topcon NW400; K240058 (April 23, 2024) added the handheld Optomed Aurora.
• Cameras: Topcon NW400 and Optomed Aurora.
• Handheld studies: Two prospective studies summarized in the 2024 clearance, with 317 and 362 participants. Sensitivity 92% and 93%, specificity 94% and 89%, imageability 99% in both.
• Worth knowing: It uses one image per eye, and it is the only option if you need portability: a satellite clinic, home visits, a mobile unit, or an exam room with no space for a tabletop camera. The handheld studies are smaller than the other vendors' pivotal trials, and the confidence intervals on sensitivity are correspondingly wide (roughly 79% to 97%).

iPredict-DR (iHealthScreen)

• Regulatory history: K253704, cleared July 2, 2026 and announced July 9. The predicate is IDx-DR.
• Camera: iCare DRSplus, a tabletop confocal fundus camera.
• Pivotal trial: 871 subjects, per the 510(k) summary. Sensitivity 85.58%, specificity 91.77%, imageability 99.08%.
• Worth knowing: This is the newest entrant. The company says it is commercially available, but there is no published real-world deployment data yet. The ADA's 2026 Standards, written before the clearance, still list three platforms.

How to read the accuracy numbers

Three cautions before you line those percentages up.

They are not head-to-head. A 2025 review in Ophthalmology Science (Teng et al.) found no studies comparing the cleared systems on the same patients. Different trials used different populations, imaging protocols, and dilation rules. A two-point gap between vendors tells you nothing.

Real-world imaging is harder than trial imaging. That same review gathered reports from health systems already using these tools. Adopters reported non-dilated gradability of 49% to 75%, well below the 96% to 99% in the trials, with sensitivity of 87% to 100% and specificity of 60% to 91%. Small pupils, cataracts, and rushed staff all cost you images. Decide in advance what happens when the camera says "insufficient quality": dilate and retry, or refer.

Positives are often wrong, and that is by design. Work it through with illustrative numbers. Screen 1,000 patients where 10% truly have mtmDR, using a system with 87% sensitivity and 90% specificity. You get 87 true positives and 90 false positives. About half of your referrals will turn out fine. A Swiss real-world study of IDx-DR (Riotto et al., Journal of Clinical Medicine, 2024; 1,141 patients) reached the same practical conclusion: negative results were trustworthy, but the system consistently overcalled disease, so every positive needs a specialist's eyes.

That tradeoff is acceptable for a screening test. It does mean you should tell patients a positive result means "you need a full eye exam," not "you have retinopathy."

What these systems do not do

• They detect diabetic retinopathy only. Digital Diagnostics' labeling says so plainly: the device "is only designed to detect diabetic retinopathy" and will not identify other eye or systemic disease. Glaucoma, macular degeneration, and cataracts are outside the clearance for all four.
• They are for patients with no prior DR diagnosis. Patients already being followed for retinopathy belong with their eye care provider.
• They will miss some disease. Every one of these trials had false negatives. A negative result is not a reason to ignore new visual symptoms.
• They are not a comprehensive eye exam, as the ADA language above makes clear.

Getting paid: CPT 92229 and HEDIS

CPT 92229 covers retinal imaging with point-of-care autonomous analysis and report. We covered its history in Does Medicare Pay for AI?. The national Medicare rate was $47.06 in 2022, $45.75 in 2023, $40.94 in 2024 and $43.67 in 2025, according to an L.E.K. Consulting analysis published this month. A 2026 rate sheet prepared by Corcoran Consulting Group lists the national figure at roughly $47. Look up your own locality before you build a budget on it; commercial and Medicaid rates vary.

Two billing details from that same rate sheet:

• The code is bilateral. Bill it once whether one or both eyes are imaged.
• NCCI edits generally prevent billing 92229 on the same day as fundus photography (92250), remote imaging (92227, 92228), or posterior-segment OCT (92133, 92134). This matters most for optometry practices that already bill those codes.

HEDIS. The Eye Exam for Patients With Diabetes (EED) measure covers members ages 18 to 75. Health plan tip sheets for measurement year 2026 list 92229 among the qualifying codes and state that results may be read "by a system that provides an artificial intelligence (AI) interpretation." For a practice in a value-based contract, closing that gap can be worth more than the fee-for-service payment. Confirm with each plan how they want the result documented.

What it costs you. None of the four vendors publish pricing. Expect a camera purchase or lease plus a software fee that is either per exam or a subscription. Ask each vendor for both numbers and work out how many exams a month you need to break even at your actual payer mix.

Why almost no one is using it

The technology has been available since 2018 and billable since 2021. Uptake is tiny.

• A JAMA Ophthalmology study (Shah et al., December 2024) used the TriNetX database of nearly 5 million patients with diabetes. Just 3,440 had an AI exam billed under 92229 between 2021 and 2023. That is 0.09%. The rate was flat: 58.0 per 100,000 in 2021 and 58.6 in 2023.
• A Medicare claims analysis in Ophthalmic Epidemiology (Satheesh et al., August 2026) counted 1,427 services under 92229 in 2023, billed by 39 providers in 16 states. Optometry accounted for 57.7% of them. Family practice was 13.9%.

The second finding is a surprise. A code built for primary care is being used mostly by optometrists.

Where it has been deployed with commitment, the results are good. At Johns Hopkins primary care sites, clinics that switched to autonomous AI saw a 7.6 percentage point greater increase in annual diabetic eye testing than clinics that did not, and adherence among Black patients rose 12.2 points (Huang et al., npj Digital Medicine, 2024). In the ACCESS randomized trial in young people with diabetes, 100% of the point-of-care AI group completed an eye exam within six months versus 22% of the group given a standard referral (Wolf et al., Nature Communications, 2024).

That same trial shows where the hard part moves. Of 25 participants with an abnormal AI result, 16 (64%) followed through with an eye care provider. Better than the control arm, but a third of the patients who most needed a specialist still did not get to one.

Which one fits your practice

• You need portability or have no room for a tabletop camera: AEYE-DS with the Optomed Aurora is the only cleared handheld pairing.
• You want a separate flag for vision-threatening disease: EyeArt.
• You already own a Topcon NW400: three of the four systems run on it, so you can compare quotes on software alone.
• You already own a Canon CR-2 AF or CR-2 Plus AF: EyeArt. If you own an iCare DRSplus: iPredict-DR.
• You want the longest track record and the most independent studies: LumineticsCore.
• You are considering iPredict-DR: ask for references from live sites. It is cleared, but it is three months into the US market.

Whichever you pick, settle these before go-live:

1. A named eye care partner who will take your positives and your ungradable patients, with an agreed turnaround.
2. A tracking list for every positive result until the specialist visit is confirmed.
3. A dilation policy for insufficient-quality images.
4. Patient eligibility rules on the order itself: no prior DR diagnosis, no visual symptoms, within the labeled age range.
5. A script for staff explaining that the test checks for diabetic retinopathy only and does not replace a full eye exam.

The general rollout habits from our AI scribe implementation guide apply here too: pick one site, name an owner, and measure before you expand.

Related reading

• AI medical imaging and diagnostics tools in the directory
• Directory pages for family medicine physicians, optometrists, and small medical clinics
• FDA's 2026 clinical decision support guidance, for where the line between device and non-device software sits
• AI for dental practices: what is FDA cleared, the closest parallel in another specialty

---

This article is for informational purposes only and is not medical, legal, or billing advice. FDA clearance details were taken from the openFDA 510(k) database and published 510(k) summaries as of September 30, 2026. Performance figures come from separate studies and are not directly comparable. Payment rates vary by payer and locality. Verify indications, supported cameras, pricing, and coverage with each vendor and your payers before purchasing.