Oatmeal Health

Three distinct model applications, all described as the company's own rather than licensed, which sets this apart from several peers on the same source list. A model over structured and unstructured record data identifies patients meeting the recommended screening criteria, with a stated area under the receiver operating characteristic curve of 0.92. A vision transformer detects and risk stratifies pulmonary nodules on low dose computed tomography.

A language model generates patient education and drives a follow up chatbot. Held at B rather than A for two reasons that pull in the same direction: a clinical team is part of what a customer buys, described alongside the technology rather than behind it, and the diagnostic model that anchors the whole proposition is not cleared, so the component doing the most consequential work is not yet the component in production use.

The imaging half is bounded correctly and the patient communication half is not described at all. On imaging the framing is consistently adjunct: a malignancy risk score presented inside the detection viewer radiologists already use, positioned as augmenting the established nodule reporting classification rather than replacing it, with the radiologist reading and reporting. That is the right posture and the company states it clearly. The gap is the other component.

A language model chatbot explains screening results and next steps directly to patients, in their preferred language, tailored to their medical history, and schedules their follow up. That is generative output about cancer screening findings delivered to a patient with no described clinician review, no published guardrails, no escalation rule for a distressed or confused patient, and no stated handling of the case where a patient asks what the result means for them.

The population served has, by the company's own account, lower health literacy and greater language barriers than average, which raises rather than lowers the standard. This is the element to ask about first and it is not the element the company foregrounds.

More specific than most of this list and still not checkable. The company names the architecture rather than gesturing at artificial intelligence: a vision transformer for nodule work, a stated discrimination figure of 0.92 for candidate identification, the named clinical criteria the eligibility model targets, and the named reporting classification the risk score augments.

Naming an architecture and a metric is a real step above the vendors in this batch whose entire description is that they are artificial intelligence driven. What is absent is everything needed to verify it: no publication, no described training or validation cohort, no external benchmark, and a claim of state of the art discriminative performance with neither a number nor a comparator attached.

One small precision point worth noting because it recurs across this market: the discrimination figure is glossed on the company's own page as accuracy, and those are different measurements, since discrimination describes ranking across all thresholds while accuracy describes correctness at one.

Nothing a third party can check. Two retrieval passes located no peer reviewed publication, no registered trial, no independent validation and no named cohort. The available figures are vendor stated and unaccompanied by denominators: a 90 percent patient conversion rate for telemedicine shared decision making visits in pilot programmes, a claim of identifying high risk patients with 94 percent accuracy, and the 0.92 discrimination figure.

Because the diagnostic component is pre clearance, there is also no regulatory review record to serve as the substitute evidence trail that a cleared device provides. The deployments themselves are real and named across health centres in several states, so the operational footprint is not in doubt; what is missing is any measurement of it that someone outside the company produced or could reproduce.

Two passes located no retention position, no minimisation statement and no secondary use policy. One question is specific to this architecture and is newly important across this whole market: the platform runs language model generated patient education and a patient facing chatbot, and nothing published states whether that generation happens on a self hosted model or through a third party model provider.

If it is the latter, patient information including medical history is leaving to a model vendor, and the relevant questions are which provider, under what agreement, with what retention, and whether the content is excluded from training. A company handling imaging, full record data including clinical notes, and patient conversations has three separate data types in play and describes the handling of none of them.

Two retrieval passes located no HIPAA statement, no Business Associate Agreement terms and no privacy or legal page. Graded on published posture. Agreements certainly exist, since federally qualified health centres are covered entities and could not have deployed otherwise.

The subprocessor point applies with particular force here because of the language model component: if generation runs through an external provider, that provider handles protected health information and must be bound as a subcontractor, and a customer should establish which entities sit in that chain rather than assuming the agreement it signed covers them.

Two passes located no SOC 2, no HITRUST, no ISO 27001, no trust centre and no vulnerability disclosure policy. Worth stating plainly rather than as a generic absence: the customers are safety net providers with limited security assessment capacity of their own, which reduces the chance that a gap gets caught during procurement and therefore raises the value of a published attestation rather than lowering it. Health centres serving low income populations are a well documented target for healthcare breaches, and they are the least equipped to run their own third party risk assessment.

Two statements about this product cannot both be true at once, and a buyer needs to see them side by side. The company's own site states that the offering is pre 510(k) with clearance expected in early 2027, which is a clear and conservative description of a device still in development. Press coverage from January and February 2025 described a fully reimbursable second opinion diagnostic tool covered by Medicare, holding one of only sixteen artificial intelligence specific procedure codes.

A device that has not been cleared cannot be marketed in the United States as a reimbursable diagnostic, and promoting an uncleared device carries its own exposure separate from whether it eventually clears. The most favourable reading available is that the earlier statements described an intended pathway, or referred to a billing route belonging to a partner product rather than to this one, and neither reading is stated anywhere located.

Graded below the ordinary status for an unregulated product precisely because the ambiguity works in the vendor's favour and against the buyer's: a health centre reading the 2025 coverage would reasonably conclude it was buying something reviewed and payable, and the current site says it is not yet either.

A genuine equity purpose, operationalised rather than merely stated, and no evidence that the models serve it. The credit is real: the company exists to reach populations that lung cancer screening has reached least, works through federally qualified health centres and safety net partners in several states, builds multilingual and culturally specific patient communication, and names in its own material that minority communities are both underserved in screening and underrepresented in the trials that establish how screening tools perform.

Very little in this index is designed around that problem from the outset. Against it, no subgroup performance is published for either model, and the omission is sharper here than it would be elsewhere, because a company whose founding argument is that existing tools underperform for underserved populations is the company most obliged to show that its own do not.

One structural point compounds this: the eligibility model is trained to identify patients meeting the established screening criteria, and those criteria are themselves documented as identifying eligible Black patients less reliably, because the smoking history thresholds interact with different exposure patterns. A model that faithfully reproduces the criteria inherits that, and faithfulness to a biased rule is not neutrality.

The most specific integration claim in this batch, and specificity is what earns the grade. Named record systems include Epic, eClinicalWorks, Athena, Azara and NextGen, which is a list weighted toward the systems community health centres actually run rather than only the large enterprise ones. On the imaging side the platform states it can read low dose scans from any major scanner manufacturer and integrate with any picture archiving system.

The company also states that no manual data cleaning or post processing is required, since the model consumes structured and unstructured record data as it finds it, which removes the preparation project that usually consumes the first months of an implementation. Held at B because all of this is the vendor's own description with no implementation detail or customer confirmation located.

A hosted platform with no published architecture, named region or residency commitment located in two passes. The imaging pathway makes the question concrete rather than abstract: reading low dose scans from any scanner and any archive means diagnostic images move somewhere for inference, and nothing states whether that is inside the health centre's environment, inside the vendor's, or inside a cloud region. The language model component raises the same question a second time for a different data type. Two data flows, both unaddressed.

No pricing, pricing mechanism or contracting model located. The company does publish an economic thesis, which is that a reimbursable procedure code attaches to every scan, and ordinarily naming the payment mechanism would earn credit on this axis. It cannot here, because the reimbursement claim sits against the company's own statement that the diagnostic component is not yet cleared, so a purchasing health centre cannot rely on it when modelling the decision.

Federally qualified health centres operate on tight margins and are precisely the buyers least able to absorb a programme whose revenue assumption does not materialise, which makes clarity on this point more important for this customer base than for almost any other.

Deliberately narrow and coherent. One cancer, one screening modality in low dose computed tomography, one buyer type in federally qualified health centres and the health plans covering the same populations, and one country. Named partnerships span health centres and primary care associations in Oklahoma, Alaska, Massachusetts and Illinois.

The narrowness is a reasonable strategy rather than a limitation, since lung cancer screening has among the lowest uptake of any recommended cancer screening and the gap is concentrated exactly where this company sells. Graded C on breadth while recording that the whole pathway from eligibility through scan through follow up is covered within that narrow scope, which is more of the journey than most single purpose vendors attempt.