Eyenuk

The algorithm is the entire product and it replaces the reader rather than assisting one. A retinal photograph goes in and a diagnosis comes out with no clinician interpreting the image at any point, which is the strongest form this axis can take.

The hardware belongs to other companies. The system is cleared for use with named cameras made by two different manufacturers, so what is sold is purely the grading, running on equipment the buyer already has or purchases separately. There is no imaging device, no reading service and no platform underneath.

This sits at the far end of the autonomy spectrum in this index, alongside the autonomous reporting radiology products, and the reason it earns an A rather than a warning is that the boundaries are defined externally and published rather than asserted.

The indication is narrow and regulator set: adults with diabetes not previously diagnosed with more than mild disease. The cameras are enumerated and each was separately cleared. The output is per eye rather than per patient, so one eye can refer while the other does not. The care pathway around it is stated, with detected cases referred to an ophthalmologist and screening performed by primary care staff who are not asked to interpret anything.

The Norwegian national deployment is the clearest expression of what that means: grading in under 30 seconds with no ophthalmologist review, at national scale, with the result written straight into the record. A regulator, a notified body and a national procurement have each examined that arrangement.

Precise where it counts for a device, which is scope rather than architecture. The clearance number is public, the indication is stated exactly, the two detected severities are named, the per eye output is described, product versions are identified individually, and camera compatibility is enumerated model by model with each addition separately cleared.

That last detail is more informative than it looks, because it tells a buyer that performance is understood to depend on the imaging device and that the vendor has not generalised beyond what it has tested. Held below A because no model architecture is described and no headline accuracy figure appears in company material, so the performance evidence sits in the regulatory file and the literature rather than in front of the reader.

Nothing identifies any party in the chain: no model or model family, no hosting arrangement and no sub processor list was located in two passes, and no published position covering retinal image retention, de identification or model improvement was found. The data is more identifying and more informative than it appears, which makes the gap specific rather than general.

A retinal photograph is a biometric image, unique to the individual and stable over time, and research has repeatedly shown that retinal images carry signal well beyond the eye, including cardiovascular and neurological information.

The company itself markets predictive biomarker work alongside screening, which converts that from a theoretical property into a live question: a patient attending for a diabetic eye screen has consented to be told whether they have retinopathy, and a single image of their retina may also be capable of supporting inferences about their heart or their brain that they neither requested nor were warned about.

Whether anything beyond the screening finding is derived from a screening image, whether such derivations are retained, and whether the patient knows, is unaddressed. Ask what is derived from a screening image beyond the cleared indication, what is retained, whether images feed the biomarker programme, and what patients are told.

Validation arrives from four independent directions, which is what separates this from records resting on a single clearance.

A pivotal trial supported the original clearance. Independent peer reviewed literature examines the system, including a feasibility study of point of care screening in general practice that reports honestly on image quality and workflow difficulties as well as on performance. A European notified body assessed it to Class IIb for three diseases, which is a separate and more demanding review than the American clearance. And Norway selected it through a competitive national procurement, meaning a public body with no interest in the outcome compared it against rivals and chose it.

Real world deployment reinforces rather than substitutes for that: more than 25 centres in Italy, a German diabetes clinic, rural Canadian screening and United States safety net clinics.

Graded on an honest basis. No published stewardship position covering retinal image retention, de identification or model improvement was located in this pass.

The data is more identifying than it appears. A retinal photograph is a biometric image, unique to the individual and stable over time, and research has repeatedly shown that retinal images carry information well beyond the eye, including cardiovascular and neurological signal. The company itself markets predictive biomarker work alongside screening, which makes the question of what else is derived from a screening image, and whether the patient knew, a live one rather than a hypothetical.

Graded on an honest basis and flagged for re verification. No compliance statement or agreement posture was located in this pass.

The jurisdictional spread makes this more complicated than a single market vendor. Operating in the United States, the European Union, Norway and Canada means four different data protection regimes apply, and the Norwegian arrangement in particular writes results into a national electronic record, which will have been governed by terms set in the procurement rather than by anything the company publishes.

Recorded honestly: the dedicated trust and security search this index requires was not run in this pass, so the grade is provisional and should not be quoted until it has been.

One consideration will bear on the eventual grade and points upward. A national health system procurement, run by a state purchasing organisation on behalf of four regional health authorities and resulting in integration with the national electronic record, involves a security and data protection assessment far more demanding than a commercial vendor review. That assessment has evidently been passed. What it produced is not published.

Among the deepest regulatory positions in the index, and unusually it is deep in three jurisdictions rather than one.

United States: 510(k) clearance K200667 in August 2020, the first autonomous system able to detect both more than mild and vision threatening disease in a single test, and the first returning a diagnostic output per eye. A later version added a third camera from a second manufacturer, making it the only cleared system not tied to a single imaging vendor.

Europe: Class IIb marking under the medical device regulation covering three diseases from one exam, diabetic retinopathy with macular edema, age related macular degeneration and glaucomatous optic nerve damage. Class IIb requires a notified body to audit the quality system and examine the technical file, which is a materially higher bar than self declaration.

Canada: approved by the national regulator. The company is the only one holding both the three disease European marking and United States clearance, and the asymmetry is worth noting on its own: the same system is authorised to detect three conditions in Europe and one in the United States.

No subgroup performance was located in company material, and the grade is C rather than lower because a pivotal trial and a notified body technical file both ordinarily contain it. Re verify against the clearance summary and the published trial.

The reason to pursue it here is specific and cuts against the usual framing. This product is deliberately deployed where screening is hardest to reach: safety net clinics, rural and remote communities, national programmes serving whole populations. That is a genuine public health good and it is precisely why unpublished subgroup performance matters more, not less. Retinal image quality varies with cataract, pupil size, media opacity and fundus pigmentation, all of which differ systematically across populations. And the consequence of a false negative in autonomous screening is distinctive: the patient is told nothing was found and does not return for a year, with no clinician having looked at the image in the meantime.

The scope discipline here is the finding, and it is more informative than it first appears. The clearance number is public, the indication is stated exactly, the two detected severities are named, the per eye output is described, product versions are identified individually, and camera compatibility is enumerated model by model with each addition separately cleared. That last item is a limitation disclosure in disguise.

Enumerating compatible cameras and clearing each addition separately tells a buyer that performance is understood to depend on the imaging device and that the vendor has not generalised beyond what it has tested, which is exactly the failure mode that breaks imaging models in the field: a system validated on one camera meets another with different optics, illumination and resolution and degrades without announcing it.

A vendor willing to say its product works with these specific devices and not others is accepting a commercial cost to avoid an unsupported claim. Held below the top grade because no model architecture is described and no headline accuracy figure appears in company material, so the performance evidence sits in the regulatory file and the literature rather than in front of the reader, and no warranty, indemnity or remediation commitment attaches. Ask for sensitivity and specificity per camera model and per severity, and for the ungradable image rate.

The Norwegian deployment demonstrates integration at a level few records here can show: results written directly into the national electronic record system used across the health service, rather than delivered to a portal for someone to transcribe.

The other integration surface is the camera, and it is handled unusually well. Clearance covers named models from two manufacturers, so the system fits imaging estates the buyer already owns instead of requiring a specific device. Held at B because no interface standard, named United States record vendor integration or application interface detail was located outside the Norwegian arrangement.

Better evidenced than most by the range of settings it demonstrably runs in: primary care clinics, eye care practices, safety net health centres, mobile and remote screening in Canada, and a national programme in Norway. Grading returns in under a minute at the point of care and under 30 seconds in the Norwegian configuration.

Operating inside a European national health service with results landing in the state electronic record implies an architecture that satisfies European residency requirements, which a United States only cloud service would not. The specific hosting model, regional arrangements and image retention schedule are not published, which is what holds this below A.

Stronger than almost anything in this index that does not publish a price, because the payment mechanism is public policy rather than a vendor decision.

A dedicated procedure code exists for artificial intelligence based diabetic retinopathy screening, introduced in 2021 as the first code specific to artificial intelligence, and it permits a primary care practice to bill for the screening without specialist oversight. This system is among those covered under it. That means a buyer can establish what the activity reimburses from public fee schedules rather than from the vendor, and can model the economics before ever speaking to a salesperson, which is exactly the transparency this axis exists to reward.

Held at B because the company publishes no licence price, no per screen rate and no indication of how its own charge relates to the reimbursement, so the margin between what a clinic is paid and what it pays remains the vendor's to disclose and it has not.

Narrow in disease and wide in setting, which is the point of the product. The clinical scope is diabetic retinopathy in the United States and three retinal conditions in Europe, and nothing beyond the eye.

Where it reaches is the interesting half. Because no specialist is required, screening can happen wherever a camera and a member of staff can be put: primary care, endocrinology, safety net clinics, rural and remote communities, and a whole national programme. Geographic coverage spans the United States, European Union, Norway and Canada, which is broader than most records here. The buyer ranges from a single primary care practice to a state purchasing authority, and those are very different sales.