Digital Diagnostics

The model is the product, without qualification. LumineticsCore analyses retinal images and issues a diagnostic result for more than mild diabetic retinopathy with no physician interpretation of the image at any point. Remove the algorithm and nothing remains but a camera.

This is the cleanest centrality A available: the device exists because an FDA authorisation was granted for the algorithm itself, and the regulatory record describes the algorithm as the subject of the authorisation rather than as a feature of a platform.

The furthest point on this index's autonomy spectrum and the only United States authorised instance of it. The system returns a diagnosis to the patient at the point of care without a clinician reading the image. The grade is an A not because the autonomy is maximal but because the oversight model around it is explicitly constructed rather than left implicit.

Autonomy is bounded by a narrow authorised indication covering adults aged 22 and over who carry a diabetes diagnosis and have not previously been diagnosed with diabetic retinopathy, with defined behaviour outside that boundary and a defined referral action on a positive result.

The company states that the creator of the autonomous AI assumes liability for the diagnosis, which resolves the accountability question that almost every other autonomous or semi autonomous product in this index leaves open. Compare Ibex, which deliberately places its read after the pathologist to eliminate automation bias. Both are considered positions on where the human belongs, which is what this axis rewards.

Substantially more inspectable than the norm, and the transparency is on the public regulatory and scientific record rather than in marketing. The De Novo authorisation established special controls that now bind every subsequent device clearing into this category, and the device description on the 510(k) record covers the guided acquisition workflow, DICOM image submission and local image retention added in the current version.

The pivotal trial was preregistered, so the endpoints were fixed before results were known. The founder has published on the system's design and its ethical framework in peer reviewed venues. Held at B rather than A because the model architecture, training data composition and weights are not published, so an outside party cannot independently characterise where performance degrades. Knowability here comes from a regulator having reviewed the submission, not from the public being able to inspect the model.

The architecture is favourable and the published detail is thin. The device operates at the point of care and the current version added local image retention, both of which limit how far retinal images travel and are structural properties rather than policy promises, so for the acquisition and reading path a buyer can reason about the exposure without needing a disclosure statement. What is absent is everything downstream of that.

Two retrieval passes located no published statement covering retention periods, deletion rights, secondary use of images for model improvement, or whether patient images contribute to training, and no hosting arrangement or sub processor list was named. The input type is what makes the secondary use gap the specific one rather than a general omission.

A retinal photograph is biometric data, unique to the individual and stable across a lifetime, with identification potential beyond ordinary clinical imagery, and it is captured here in a primary care office from patients attending for a screening test rather than a specialist consultation. A biometric image taken in that setting carries a narrower expectation about what it will be used for than one taken in an ophthalmology clinic. Ask what is retained locally and for how long, whether anything leaves the device, whether images contribute to model development, and what the patient is told at the point of capture.

The strongest evidentiary design in this index and the reason the grade reaches A. The pivotal study was prospective, preregistered and run at primary care sites, which is the design the index has repeatedly said earns an A and which comparable vendors have not produced. Reported performance is roughly 87 percent sensitivity and 90 percent specificity for referable disease. Equity was tested as part of the pivotal work rather than examined afterwards.

The SEE study, published in Diabetes Care in 2021, extended safety, efficacy and equity testing to a youth population. Independent literature exists beyond the company: a review in Ophthalmology Science assessed LumineticsCore alongside EyeArt and AEYE Diagnostic Screening and gathered implementation experience from academic health systems, which places this vendor inside a third party comparison rather than only in its own studies.

One conflict is recorded plainly rather than treated as disqualifying: the founder is an author on much of the pivotal and framework literature, disclosed in those publications. The same disclosure discipline is applied to David Bates in the MedAware record and to the Lyssn founders.

The architecture is favourable and the published detail is thin. The device operates at the point of care and the current version added local image retention, both of which limit how far retinal images travel, and retinal images are biometric data with identification potential beyond ordinary clinical imagery.

Two retrieval passes located no published statement covering retention periods, deletion rights, secondary use of images for model improvement, or whether patient images contribute to training. For a product whose entire input is a biometric image captured in a primary care office, the absence of a published position on secondary use is the specific gap, and it is the question a buyer should ask first.

Graded on published posture. Two retrieval passes located no HIPAA statement, no Business Associate Agreement terms and no privacy or legal page from the company itself. A third party directory describes the product as HIPAA compliant, which is not corroborated in vendor material and is not graded on, consistent with the source hygiene applied to VigiLanz and FeelBetter.

A Business Associate Agreement certainly exists contractually or the deployments and the Medicare billing would be impossible, so this grade describes what a buyer can read before contracting rather than what is contractually true. This is the axis most likely to move on a direct request to the company.

Two retrieval passes located no SOC 2, no HITRUST, no ISO 27001, no trust centre and no vulnerability disclosure policy. An ISO 13485 quality management certification is plausible for a company holding a Class II device authorisation but none was located, so none is asserted.

The shape of this grid is itself the finding and it repeats a pattern the medication safety lane established: the vendor that did the hardest regulatory and evidentiary work in its category publishes the least about how it secures the system. Regulatory rigour and security publication are independent, and buyers routinely read the first as evidence of the second.

The strongest regulatory position in this index. De Novo authorisation granted April 2018 as a Class II device, following expedited review under the Breakthrough Devices programme, making it the first autonomous AI diagnostic system authorised by the FDA in any field of medicine. A 510(k) for the current version cleared in May 2021 using the original device as predicate, with the De Novo special controls carried forward.

The authorisation did not merely clear a product, it created the regulatory category, and competitors including EyeArt subsequently cleared against LumineticsCore as predicate. Two distinctions a buyer should hold: the authorised indication covers adults aged 22 and over only, so paediatric use is outside it despite published paediatric evidence, and De Novo authorisation is a substantively different and more demanding route than a 510(k) that relies on an existing predicate.

Equity was a design input to the pivotal evidence rather than a disclosure added afterwards, which is what separates this record from almost every other in the index. The pivotal work assessed equity alongside safety and efficacy, and the SEE study in Diabetes Care carried the same three part framing into a youth population.

Beyond its own product, the founder co authored the npj Digital Medicine work on addressing bias in artificial intelligence for health equity and a peer reviewed reimbursement framework that sets out the ethical guardrails built into the system's implementation. The liability position reinforces this: a company that accepts responsibility for a diagnosis its model issues has a direct incentive to characterise where that model fails.

Comparable in strength to the Cognivue FOCUS study, which is the index's existing governance benchmark. One item is owed and is recorded rather than assumed: subgroup performance figures for the authorised adult indication broken out by race, ethnicity or retinal pigmentation were not located in this assessment, and pigmentation is a plausible source of variation in retinal image analysis.

One methodological choice does most of the work here and it is the strongest single guard against a favourable result being manufactured. The pivotal trial was preregistered, so the endpoints were fixed before the results were known. That removes outcome switching, which is the quiet mechanism by which a study that missed its primary endpoint gets reported on a secondary one that happened to work, and it is why a preregistered trial carries more weight than a larger unregistered one.

Around it, the authorisation route established special controls that now bind every subsequent device clearing into this category, so a public standard exists that the company did not write and competitors must meet. The device description on the regulatory record covers the guided acquisition workflow and local image retention, and the founder has published on the system's design and ethical framework in peer reviewed venues.

Held below the top grade because model architecture, training data composition and weights are unpublished, so an outside party cannot independently characterise where performance degrades, and no warranty, indemnity or remediation commitment attaches. The distinction is worth stating plainly: knowability here comes from a regulator having reviewed the submission, not from the public being able to inspect the model. Ask for performance by patient subgroup and image quality band, and what the system does when acquisition is inadequate.

Confirmed from the regulatory record: the current version accepts DICOM image submission, which is the standard that matters for imaging workflow. Not confirmed and therefore not graded on: a third party directory describes report delivery into Epic, pairing with Topcon cameras, and deployments at OCHIN sites and Labcorp centres.

None of that was corroborated in company material during this assessment, and the index has been burned by exactly this pattern before, most recently on PrecisePK where a machine learning claim appeared only on a third party page. Graded C on what is verifiable.

The camera dependency is worth naming for buyers regardless of which manufacturer: an autonomous diagnostic system that requires a specific class of nonmydriatic fundus camera carries a hardware purchase and a hardware relationship that a software licence discussion will not surface.

Deployed at the point of care in the clinic where the image is captured, and the current device version added local image retention, both established on the regulatory record rather than in marketing. That is the favourable answer to the question this index asks across categories, which is where the data goes: the cardiovascular CT lane splits on where the scan goes and the drug discovery lane splits on where the chemistry goes. Here the image can stay where it was taken.

Held at B rather than A because no published data residency statement, hosting region or architecture description was located, so the residency posture is inferred from the device description rather than stated by the company.

Unusual and worth naming as a candidate benchmark. CPT code 92229 was established for remote retinal imaging with a national Medicare payment amount finalised in the calendar year 2022 Physician Fee Schedule and Outpatient Prospective Payment System rules, the first national payment established for an FDA authorised autonomous AI service. So a buyer can establish the reimbursement side of the economics from public rulemaking rather than from the vendor.

More unusual still, the founder published a peer reviewed reimbursement framework in npj Digital Medicine setting out how value and cost for the service were analytically determined and which stakeholders were involved. Publishing an account of how your own service came to be priced is something no other vendor in this index has done.

Held at B because the company's own charge to the provider is not published, so the provider still cannot compute margin without entering a sales conversation. Artrya remains the A on this axis because a stock exchange listing forces disclosure of the price as well as the mechanism.

Deliberately narrow, and the narrowness is a feature rather than a deficiency. The authorised indication covers one finding, more than mild diabetic retinopathy including macular edema at ETDRS level 35 or higher, in adults aged 22 and over who have diabetes and no prior diabetic retinopathy diagnosis.

Setting is primary care and other point of care sites where patients with diabetes are already being seen, which is the whole strategic point, since screening rates in the United States have been reported anywhere from 11 to 70 percent and the exam usually requires a separate visit to an eye specialist. A second product, DermSpot, addresses skin, but its regulatory status was not established in this assessment and it is not graded here. Graded C on breadth while recording that a precisely bounded indication is the correct posture for an autonomous diagnostic device and should not be read as a limitation of quality.