Lucem Health

The models belong to other companies. Three Reveal programmes were developed with Medial EarlySign, whose predictive models are that company's own product and commercial asset, and the arrhythmia programme runs on an iRhythm relationship. Lucem builds the deployment: data plumbing, workflow placement, patient outreach, care programme design and measurement, which it calls AI SolutionOps. This record was built rather than rejected, and the reasoning is recorded so it can be reversed.

The Vim rejection ground covers middleware whose marketplace hosts other companies' models and leaves the health system to choose; this is a step further along, because Lucem sells a named product with a defined clinical purpose, and the buyer contracts with Lucem rather than with the model developer. The scoping rule this index uses is to index the artificial intelligence product a buyer can engage, and a buyer engages this one.

Candour is credited in the grade rather than penalised: the company describes itself accurately as working with artificial intelligence innovators, and its founding chief executive has said the hard part is integrating tools that already work rather than building them. That places it on the honest non inflation roster alongside Candid Health and Schrodinger.

A model produces a ranked list of patients who may warrant a test they have not had, and humans do everything after that: clinicians review, outreach staff contact the patient, and the diagnostic decision belongs to the ordering physician. Nothing is diagnosed and no order is placed automatically. The design is deliberately additive to existing population health processes rather than replacing them, which the company states as a selling point.

The oversight gap is structural rather than clinical and belongs to the arrangement rather than the software: a model surfacing patients and a separate organisation acting on them means responsibility for a missed patient is split across at least three parties, and no published material assigns it.

The transparency here is inherited, and inherited transparency has a gap in it. Lucem points to partner algorithms validated in multiple health systems with findings published in leading journals, which is true and is a real strength of the underlying models, but that literature belongs to the developer and describes the model in general rather than the instance a given customer runs.

What is not published by anyone is the part Lucem controls: which model version is deployed at a site, whether it is recalibrated against local data, what thresholds are set and by whom, and how those choices change who appears on the list. A published performance figure for a model is not a published performance figure for a deployment of that model, and the difference lives exactly where this company sits.

Real programmes running at real health systems, with the evidence pointing at the wrong thing. Everything cited belongs to the model developers: peer reviewed validations of the underlying algorithms across multiple health systems, and third party analyst recognition of Medial EarlySign. Those establish that the models detect what they claim to detect. They say nothing about the layer this company sells.

Lucem's entire proposition is that clinical artificial intelligence fails on implementation rather than on accuracy, so the numbers that would prove its value are implementation numbers: what share of surfaced patients were contacted, what share completed the test, how that compares to the site's baseline, and how much of the effect survives after the launch period.

None was located, and a claim of finding up to eight times the cancers in colorectal screening is stated without the denominator or the comparison it implies. The company measuring the gap it exists to close would be the single most persuasive thing it could publish.

Two passes located no retention position, minimisation statement or secondary use policy. The specific question this arrangement raises is a chain of custody one, and it is sharper than the usual version: patient data originates in a health system, moves through Lucem's platform, and is scored by a model belonging to a third company headquartered in another country.

Nothing published describes which party holds identifiable data at which step, whether the model developer receives records or only returns scores, whether anything is retained for model improvement, or which of the three parties a patient would approach to exercise a right. Three party data flows are common in this market and are almost never documented, and a buyer should ask for the map rather than the assurance.

Two retrieval passes located no HIPAA statement, no Business Associate Agreement terms and no privacy or legal page. Graded on published posture. The subcontractor question makes this more consequential than a missing page usually is: an agreement between a health system and Lucem does not by itself cover a model developer acting further down the chain, and HIPAA requires those obligations to flow down to subcontractors handling protected health information.

A health system signing here should establish in writing which entities sit in the chain and confirm that each is bound, rather than assuming a single agreement with the party it can see covers the party it cannot.

Two passes located no SOC 2, no HITRUST, no ISO 27001, no trust centre and no vulnerability disclosure policy. The absence is notable for a company whose investors and early customers include two of the larger health systems in the United States, since organisations of that size normally require attestations as a condition of contracting, which suggests the documentation exists privately and is simply not published. Publishing it would be a competitive advantage in a category where, as this index has repeatedly found, almost nobody does.

No clearance, authorisation or submission located, and none claimed. The programmes operate in the space the Cures Act leaves outside device regulation for clinical decision support, which is the ordinary position for risk stratification over record data. The certification non transfer point applies here in its supplier chain form, which this index has now recorded several times: a regulatory status held by a component does not travel to the product assembled from it.

The iRhythm relationship involves a cleared cardiac monitoring service, and that clearance covers the monitor rather than the model that decides which patients should wear one. A buyer encountering a cleared device inside a programme should establish precisely which element the clearance covers.

The structure concentrates every governance question at exactly the point where nothing is published. Lucem runs models it did not build, over data it does not own, to reach patients it never sees. That arrangement raises a specific set of questions, none of which is answered anywhere located: who validates a model against a site's own population before it goes live, who monitors for drift afterwards, who notices if performance degrades in a subgroup, and which of the three parties is accountable when it does.

No subgroup performance, bias testing methodology, model card or drift monitoring concept was located from the deploying party. This is the accountability question this index first recorded against grounding suppliers, appearing again one layer closer to the patient: the party best placed to detect a problem is the one operating the deployment, and the party that understands the model is the one that built it, and they are different companies.

This is the actual product and the company is straightforward about it. The programmes are designed to run on data a health system already has, without new instrumentation, new data collection or changes to clinical workflow, and to sit alongside existing population health tooling rather than displacing it.

Deployment speed and low integration burden are the stated value proposition rather than a footnote, and the company has stood these programmes up across multiple health systems including its own investors. Held at B rather than A because no named record system, interoperability standard or implementation detail was located, so the depth is asserted rather than demonstrated, and because low integration burden is a claim every vendor in this category makes.

No published hosting architecture, named region or residency commitment located in two passes. The question is not routine for this vendor because of the supplier chain: the models behind several programmes come from a company headquartered in Israel, so whether inference happens inside the customer's environment, inside Lucem's, or inside the developer's determines whether patient data crosses a border at all.

Those are three materially different propositions and nothing published distinguishes them. Compare Carenostics in this same index, which answers the equivalent question by stating that data never leaves the hospital, which is a straightforward commitment available to any vendor willing to make it.

No pricing, pricing mechanism or contracting model located, and the structure raises a question ordinary licensing does not: whether the model developer is paid separately, and whether a health system that later wants to run the same model directly can do so. Funding is a 7.7 million dollar round in May 2023 led by Mayo Clinic, Granger Management and Mercy, with no subsequent round located.

That is modest for a company operating clinical programmes with outreach staffing across multiple health systems, and supplier continuity deserves the attention the Behold.ai precedent established, with the added wrinkle that a programme depends on two suppliers rather than one, so it can fail at either end.

Broad for a company of this size, and the breadth is a direct consequence of the business model. Named programmes cover lung cancer, colorectal cancer, liver disease, arrhythmias, type 1 diabetes, progression from prediabetes to diabetes, and lower gastrointestinal disorders, which is seven disease areas from a small team.

Because each programme is a partner model wrapped in a common operating layer, adding a condition is a partnership and a configuration rather than a research programme, so coverage scales in a way that a company building its own models cannot match. The corresponding limit is that depth in any one area belongs to whoever built that model. Setting is United States health systems, spanning ambulatory and primary care wherever the record reaches.