Carenostics

The model is the entire product. There is no rules library, no curated content asset and no services layer underneath it: the platform reads a health system's own electronic health record data, structured and unstructured, and predicts which patients already have a disease that has never been coded. Remove the model and nothing is left.

The team is machine learning native rather than a business that added a model later, which is a distinction this index has had to draw repeatedly in the other direction. The founder ran the data mining group at Siemens Corporate Research and led artificial intelligence and health analytics at Siemens Healthcare before Deloitte and KPMG, and holds the SIGKDD Service Award, the highest service award in the field of data mining and knowledge discovery. That is context for the credibility of the mechanism claim, not evidence that the product works, which is graded separately and lower.

Conventional and appropriate. The platform surfaces a candidate patient and prompts a clinician inside the existing workflow; the clinician decides whether to investigate, and the diagnosis remains theirs. Two claims sit adjacent to the alert burden question that the medication safety lane established as the metric this kind of product turns on: a false positive rate stated as five times lower than current care practices, and clinician activation rates stated as five times higher than industry benchmarks.

Both are the right things to measure and both are vendor generated, qualified with up to, and published without a denominator, so they are recorded rather than graded on. A vendor measuring activation at all is ahead of the segment.

More methodological seriousness than the category norm and less published detail than the grade would need to rise. The company states that its work has been accepted at NeurIPS and AMGA, which would place its methods in a machine learning venue rather than only in trade press, but no citation was given and the papers were not located in two passes, so the claim is recorded and not graded on.

No model card, no architecture description, no feature specification and no published validation methodology were located. The per health system tailoring is described as a differentiator but never specified: what is retrained, what is held fixed, and how much local data a site needs before its model is fit for use are all unstated, and those are the questions a health system's own data science team would ask first.

Real deployment with unusually good sourcing, and evidence that has not yet been designed to test the product. The headline figures are vendor generated and qualified with up to: three times more chronic kidney disease patients identified and twenty times more severe and uncontrolled asthma patients at a five times lower false positive rate.

What lifts this above vendor claims alone is that Hackensack Meridian Health confirms the partnership through its own newsroom and shared a Bio-IT World Innovative Practice Award with the company, the same customer side corroboration that strengthened the MedAware record. That same newsroom also supplies the limiting detail: the results are described as retrospective experiments demonstrating an ability to identify more than half of undiagnosed stage 3 or later kidney disease.

Retrospective identification is not a measure of what happens when the prompts actually fire, which is the distinction that held MedAware at B. No peer reviewed clinical publication was located, and the deployment evidence rests essentially on one health system.

The strongest stated posture in this part of the index and the reason is architectural rather than contractual. The company states that no patient information ever leaves the hospital system, with models trained and run on the health system's own data in place. For a product whose input is the whole record including unstructured clinical notes, not moving the data at all is the most protective answer available, and it removes an entire class of risk rather than mitigating it.

Held at B rather than A because one question is unaddressed everywhere it was looked for: if a distinct model is fitted for each health system, it is not stated whether trained parameters, performance telemetry or error analyses leave the site. Patient records staying put and model artefacts staying put are different commitments, and only the first is made. That question is new and should now be asked of any vendor claiming local training.

Graded on published posture, with the scope of the assessment stated plainly. Two retrieval passes located no HIPAA statement, no Business Associate Agreement terms and no privacy or legal page, and the company's own website was not directly retrieved during this assessment, so this is the axis most likely to move on a further pass or a direct request.

A Business Associate Agreement certainly exists contractually, because a deployment inside an eighteen hospital network handling six million patient records would otherwise be impossible. The architecture arguably reduces how much a BAA has to carry, since the data does not move, but that makes publishing the terms easier rather than less necessary.

Two retrieval passes located no SOC 2, no HITRUST, no ISO 27001, no trust centre and no vulnerability disclosure policy, and the company website was not directly retrieved this pass, so the scope is stated rather than assumed. The absence is worth naming precisely because of the deployment model: software that runs inside the hospital network and reads the entire record, including notes, is a higher trust position than a vendor hosted service, not a lower one.

Running on the customer's own infrastructure is often treated in sales conversations as removing the need for attestation, when it in fact raises the question of what the software does once it is inside the perimeter.

No FDA clearance, authorisation or submission located, and none is claimed. The product operates in the space the Cures Act carved out for non device clinical decision support, where software that supports a clinician who can independently review the basis of the recommendation falls outside device regulation.

The tension a buyer should see is the same one the index recorded against NarxCare: the exclusion turns on independent reviewability, and a model that infers undiagnosed kidney disease from a whole record including unstructured notes does not produce a basis a clinician can reconstruct in their head. Nothing here is irregular, and this is the ordinary status of its whole segment. It is recorded because the segment's regulatory footing rests on a test that its technology makes harder to satisfy each year.

Equity is a stated purpose rather than a measured property. The company frames its mission around the underdiagnosis, undertreatment and health inequities of chronic disease, and the framing is well founded, since undiagnosed chronic kidney disease falls hardest on populations with the least contact with specialist care. But no subgroup performance, bias testing methodology or drift monitoring was located.

The per health system model design creates a governance question this index has not had to ask before and it should now be asked routinely: if every customer runs a differently fitted model, then every deployment is a separate model with separate performance characteristics, and no published material describes how a site's model is validated before it goes live, who signs it off, or how performance is monitored afterwards. A single vendor validation cannot cover a fleet of site specific models.

Deep rather than wide. The platform consumes structured fields and unstructured clinical notes from the health system's own record and returns prompts into the clinician's existing workflow rather than into a separate dashboard, which is the integration depth that determines whether a case finding product is used at all. Demonstrated at scale in one network of roughly eighteen hospitals holding six million patient records, with a Veterans Affairs relationship also on the record.

Held at B because breadth is unproven: no list of supported record systems was located, and the local deployment model means each new customer is an integration project rather than a configuration, which is the cost side of the same architecture that earns the data residency grade.

The clearest answer in this index to the question it keeps asking across categories, which is where the data goes. The cardiovascular lane splits on where the scan goes and the drug discovery lane splits on where the chemistry goes; here the answer is that it goes nowhere. Models are fitted and run on the health system's own electronic health record data inside the health system, and the company states plainly that no patient information ever leaves it.

That is a stated architectural commitment repeated in company and investor material rather than an inference from a device description, which is what separates this A from the B recorded on Digital Diagnostics. The residency question is closed by construction: there is no hosting region to negotiate, no cross border transfer analysis and no vendor side copy of the record to worry about at contract renewal or at acquisition.

Nothing published. No pricing, no pricing mechanism, no contracting model, no stated basis of charge and no return on investment figure. Funding is the only visible commercial fact: a five million dollar seed round led by M13 in September 2023 with Greatpoint Ventures, Gaingels and the chairman of Humana participating, plus strategic relationships with AstraZeneca, the Veterans Affairs system and Bayer G4A.

No subsequent round was located, which for a company deploying site specific models inside large health systems is worth a buyer's attention on supplier continuity grounds, following the Behold.ai precedent. A per site deployment model is expensive to run and the funding on record is modest against it.

Narrow today with a credible expansion path. Chronic kidney disease is the proven indication, with severe and uncontrolled asthma also deployed and chronic obstructive pulmonary disease and heart failure named as next. A separate line of work supports transplant centre operations. Setting is the health system, spanning inpatient and ambulatory wherever the record reaches, and the market is the United States only.

The constraint on breadth is the same architecture that produces the strengths elsewhere on this grid: each disease needs its own model and each customer needs its own fitting, so coverage grows by project rather than by release.