Eyenuk

The model is the product and it replaces the reader rather than assisting one. A retinal photograph goes in and a result for each eye comes out, with no clinician interpreting the image at any point. The K223357 clearance summary describes the software as an ensemble of clinically aligned machine learning algorithms, including deep learning, running on remote servers.

The hardware belongs to other companies. The cleared version works with named Canon and Topcon retinal cameras that the buyer supplies, so what is sold is the grading itself, with no imaging device or reading service underneath.

Fully autonomous, inside a boundary the regulator set and the company publishes. The indication is adults with diabetes not previously diagnosed with more than mild diabetic retinopathy, and the cleared cameras are named model by model. The output is per eye, negative, positive or ungradable for each of more than mild and vision threatening disease, so one eye can be referred while the other is not, and an image too poor to read says so rather than returning a negative. The operator gets image quality feedback at capture, and positive results route to an eye care professional.

Screening is performed by primary care staff who are not asked to interpret anything. The company reports a national deployment in Norway in which grading runs without ophthalmologist review and the result is written to the national record.

Precise on scope and partial on the model. The K223357 clearance summary describes an ensemble of clinically aligned machine learning algorithms, including deep learning, with image quality assessment, running on remote servers, and names the indication, the two severities detected, the per eye outputs and each cleared camera.

Camera compatibility is enumerated model by model, with each addition cleared, which shows that performance is understood to depend on the imaging device and that the company has not generalized beyond what it tested. Nothing describes the architecture, the training data or any version history beyond the cleared versions.

A short chain on the filings. EyeArt is the company's own cleared model: K200667 and K223357 are filed by Eyenuk, and the K223357 summary describes its own ensemble of machine learning algorithms rather than a licensed or third party model.

The analysis runs on remote computers in a secure data center whose operator is not named, and no subprocessor list is published.

Prospective, multicenter and graded against an independent reading center. The K223357 clearance summary reports the EN-01b study at six sites, 171 participants and 336 eyes, with the Wisconsin Reading Center as reference, on each cleared camera. A retrospective analysis of 655 participants and 1,310 eyes reports sensitivity of 95.1 percent and specificity of 86.1 percent for more than mild disease, and 95.5 and 92.4 percent for vision threatening disease.

Independent peer reviewed literature examines the system, including a feasibility study of point of care screening in general practice that reports image quality and workflow difficulties alongside performance. Deployments the company reports include centers in Italy, a German diabetes clinic, remote screening in Canada, federally qualified health centers in Delaware, and Norway's national screening program, selected through public procurement.

Retinal images leave the clinic. The K223357 clearance summary describes images sent to remote computers in a secure data center for analysis, and the company states HIPAA and GDPR compliance, but the data center's location, the retention period for images, de identification, and whether screening images are used to train or improve models are not published. The privacy policy covers the website only.

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 shown that retinal images carry signal beyond the eye, including cardiovascular and neurological information. Whether anything beyond the screening result is derived from a screening image, and whether patients are told, is unaddressed.

The company states HIPAA compliance on its home page. No business associate agreement terms, HIPAA page or notice of privacy practices is published, and the privacy policy covers the website only, so the compliance claim has no published terms behind it.

Four data protection regimes apply across the markets the company reports serving, the United States, the European Union, Norway and Canada, and the company states GDPR compliance alongside HIPAA.

SOC 2 Type 2 attestation is stated on the home page, alongside HIPAA and GDPR compliance and ISO 13485:2016 certification, which is a quality system standard rather than a security one.

There is no trust center, report access route or vulnerability disclosure policy, so the attestation is named without the artifact.

Two FDA 510(k) clearances under the retinal diagnostic software classification (product code PIB, 21 CFR 886.1100): K200667 for EyeArt in August 2020, and K223357 for EyeArt v2.2.0 in June 2023, with K200667 as its predicate. The indication is adults with diabetes not previously diagnosed with more than mild diabetic retinopathy, for detecting more than mild and vision threatening disease, on the Canon CR-2 AF, Canon CR-2 Plus AF and Topcon NW400 cameras. The company describes EyeArt as the first and only FDA 510(k) cleared autonomous AI that detects vision threatening diabetic retinopathy.

The company also reports Class IIb CE marking under the European medical device regulation for version 3.0, covering diabetic retinopathy with macular edema, age related macular degeneration and glaucomatous optic nerve damage, approval from Health Canada, and ISO 13485:2016 certification under MDSAP. The United States indication is narrower than the European one: one disease against three.

The K223357 clearance summary reports results under the operating conditions that most affect retinal image quality. In the prospective six site study, sensitivity for more than mild diabetic retinopathy was 95.9 percent on the Canon cameras and 94.4 percent on the Topcon, with specificity of 86.4 and 91.1 percent, and about 90 percent of eyes were gradable without pupil dilation against about 99 percent with it. The undilated eye, where an autonomous screen is most likely to fail, therefore has a published figure.

The summary gives the study population's makeup, 86 percent White and 10.5 percent Black, but no results by race, ethnicity, age or sex. The gap matters more here than for an assistive tool. EyeArt is sold for screening in primary care, safety net clinics and remote programs, and a false negative in autonomous screening means the patient is told the result is clear and is not seen again for a year, with no clinician having looked at the image. Retinal image quality varies with cataract, pupil size and fundus pigmentation, which differ across populations.

The company's own clearance filing carries its performance figures with their denominators. The K223357 summary reports a prospective study at six sites, 171 participants and 336 eyes, graded against the Wisconsin Reading Center: for more than mild diabetic retinopathy, sensitivity 95.9 percent and specificity 86.4 percent on the Canon cameras and 94.4 and 91.1 percent on the Topcon; for vision threatening disease, sensitivity 96.8 percent on both and specificity 91.7 and 91.6 percent. A retrospective study of 655 participants and 1,310 eyes reports 95.1 and 86.1 percent for more than mild disease. The indication, the three cleared cameras and the ungradable output bound where those figures hold, and the home page restates them as 96 percent sensitivity and 88 percent specificity.

No warranty, indemnity, error reporting route or remediation commitment is published.

The company reports that in Norway's national screening program results are written directly into the national electronic record rather than delivered to a portal for someone to transcribe.

The other integration surface is the camera. The cleared version works with named models from two manufacturers, so the system fits cameras a buyer may already own instead of requiring one device. Outside the Norwegian arrangement, the company publishes no interface standard, United States record system integration or application interface.

A single hosted service. Images are captured in the clinic and sent to remote computers in a secure data center for analysis, per the K223357 clearance summary, with results returned at the point of care in under a minute according to the company.

Where that data center is, who operates it, whether regional hosting is offered to European or Canadian customers, and how long images are kept are not published. The service runs in primary care, eye care practices, safety net clinics, remote screening programs and a national program in Norway.

The payer's side is public and the vendor's is not. The company states that EyeArt screening is billed under CPT code 92229, so a buyer can look up what the service reimburses from public fee schedules. That is the payer's rate, not a price.

No license fee, per screen charge or subscription is published, and the camera, which must be one of the cleared models, is bought separately.

Narrow in disease and wide in setting. The United States indication is diabetic retinopathy in adults with diabetes, and the company reports European marking for three retinal conditions. Because no specialist reads the image, screening can run wherever a cleared camera and trained staff are: primary care, endocrinology, safety net clinics, remote screening in Canada and a national program in Norway.

The validation behind the clearance is a six site prospective study and a retrospective analysis. The coverage the company names, from single practices to a national program, reaches beyond what that study describes.