Oatmeal Health

Three model applications of the company's own, inside a program that carries value beyond them. LungIQ uses a large language model and a classifier to find patients who meet the USPSTF screening criteria from FQHC clinical notes, where pack year history is rarely in a structured field. LungAI is a vision transformer the company trained to score nodule malignancy on low dose CT, currently investigational. Without the models, LungIQ reverts to manual chart review and LungAI does not exist.

The rest of the offer stands on its own. The consortium model connects FQHCs that lack CT scanners with partner imaging centers, builds shared decision making visits and G0296 coding into the workflow, and supports UDS reporting; with the models removed, a referral network and a screening program workflow would remain.

The imaging half is bounded correctly and the identification half routes to people; patient communication is the open question. LungAI is framed consistently as adjunct: a malignancy score inside the viewer radiologists already use, augmenting Lung-RADS, with the radiologist reading and reporting. LungIQ produces a ranked worklist with reasons, including an incomplete history band and a needs review band, that FQHC staff act on.

What is missing is the threshold and the patient side. Nothing says at what probability a patient enters the contact list or how the review bands are set. The company once described a language model chatbot explaining screening results and next steps to patients in their own language; the current site does not mention it, and if it is still offered, it is generative output about cancer screening delivered to patients with lower average health literacy and no published review, guardrails or escalation rule.

The approach is named with specific architecture and training data, without the version and update discipline behind it. LungAI is described as a vision transformer of about 100 million parameters trained on National Lung Screening Trial data, analyzing nodule morphology, whole scan context and changes between current and prior exams, and returning a continuous 0 to 100 malignancy probability that augments Lung-RADS categories rather than replacing them. LungIQ combines a large language model that extracts smoking history and risk factors from clinical notes with a classifier, and ranks patients by probability of USPSTF eligibility. Performance is stated for both (0.96 area under the curve each, on held out NLST data and a 35,000 patient FQHC panel).

What holds it below the top: no version or update practice is published, the language model is not identified, and the figures are area under the curve rather than sensitivity and positive predictive value at the operating threshold, which is what a clinician relying on the score needs.

The architecture is described and the language model provider is not named. LungAI is a 100 million parameter vision transformer the company says it trained on National Lung Screening Trial data, delivering a malignancy score inside the radiologist's existing CADe viewer through imaging center CADe partners, with Coreline Soft named as one. LungIQ reads FQHC record exports with a pipeline in which a large language model extracts smoking status, pack years and tobacco related risk factors from clinical notes and a classifier combines them. The published business associate agreement binds subcontractors in general terms.

The gap is the model layer the company does not build. Which large language model reads FQHC clinical notes, whether it is self hosted or reached through a provider, and what that provider retains are not stated on the home, FQHC, imaging or business associate pages, and no hosting provider or subprocessor list is published. Clinical notes are the most revealing data in the record, and the patients are safety net patients with little practical recourse.

Vendor reported figures with the validation method named and no report behind them. LungIQ is described as reaching 0.96 area under the curve for predicting USPSTF eligibility, validated on a 35,000 patient FQHC panel, with 94 percent precision among the top ranked patients contacted. LungAI is described as reaching 0.96 nodule level area under the curve in internal retrospective validation on held out data from the National Lung Screening Trial, illustrated with four held out cases. The company itself states that prospective performance has not been established and that results may differ in clinical use.

That is the company's own measurement on a named dataset, with no published report, denominator by outcome, confidence interval or sensitivity and positive predictive value at the deployed threshold. Area under the curve describes ranking across all thresholds, not correctness at the one in use. The pivotal study at Mass General Brigham, launching in the second quarter of 2026, is where independent prospective evidence would come from. Named FQHC partnerships in several states establish the operational footprint.

Retention and deidentification are settled in the business associate agreement, and training use is left open. The published agreement requires minimum necessary use, return or destruction of protected health information at termination, and de identification only under the HIPAA criteria, after which the data falls outside the agreement. It also permits data aggregation across the company's other covered entity customers.

What is not stated is whether customer data trains the models. LungAI is described as trained on National Lung Screening Trial data, a public research dataset, but nothing says whether FQHC exports or imaging center scans are used to retrain LungIQ or LungAI, and the large language model that reads clinical notes is not identified, so what its provider retains is unknown. FQHC exports include full clinical notes, the least redactable part of a record.

Oatmeal Health publishes its business associate agreement for reading before contracting, states its business associate status, applies the agreement to every provider that uses the service, and binds subcontractors through it. The business associates page, linked from the footer, carries Oatmeal Health's click through business associate agreement: it binds any agent or subcontractor that receives protected health information to the same restrictions, requires breach and security incident reporting, limits use to the minimum necessary, and requires return or destruction of protected health information at termination. The FQHC page says record exports are sent under a signed agreement, and a separate HIPAA authorization page covers patients.

Two clauses stand out. The agreement lets Oatmeal Health de identify any protected health information under the HIPAA criteria, after which it is outside the agreement, and it permits data aggregation services across Oatmeal Health's other covered entity customers. Which subcontractors actually receive data, including the provider behind the large language model that reads clinical notes, is not published.

Nothing independent stands behind the controls. The published business associate agreement commits the company to the HIPAA security requirements and breach reporting, which is a contractual obligation rather than an attestation, and no SOC 2 report, HITRUST certification, ISO 27001 certificate, trust center or vulnerability disclosure route appears on the home, FQHC, imaging or business associate pages.

The customers raise the stakes. FQHCs send full clinical note exports as flat files, and safety net providers have limited capacity to assess a vendor's security themselves, so a published attestation matters more for them than for a well resourced health system, not less.

The regulatory position is stated plainly and consistently, and the pathway is in progress. LungAI, the nodule malignancy scoring model, is a clinical output (a malignancy probability on a named patient's scan), and the site says so accurately throughout: the home and imaging center pages call it an investigational medical device that has not been cleared by the FDA and is not commercially available in the United States, and the imaging page states that a 510(k) pre submission was filed in March 2026 with a pivotal study at Mass General Brigham launching in the second quarter of 2026. Reimbursement under CPT 0721T is marketed only as following FDA clearance. LungIQ, the eligibility model, ranks patients by probability of meeting the USPSTF screening criteria for outreach and shared decision making visits, which is population health work.

Early 2025 press described a Medicare covered diagnostic, which conflicted with the site at the time. The current site does not repeat that claim, so nothing on the vendor's own pages contradicts the stated position, and the 2025 coverage described a product that is still pre clearance.

An equity purpose put into practice, with no evidence yet that the models serve it. The company exists to reach populations lung cancer screening has reached least, works through federally qualified health centers and primary care associations in several states, and builds LungIQ to recover smoking history from clinical notes because pack years are rarely in structured fields at these centers. That is a real design choice aimed at a real gap.

No subgroup performance is published for either model, an omission that cuts against the company's own argument that existing tools underserve these populations. LungIQ is trained to find patients who meet the USPSTF criteria, and criteria built on pack year thresholds have been documented as identifying eligible Black patients less reliably; a model that reproduces the criteria faithfully inherits that. LungAI is trained and validated on National Lung Screening Trial data, a cohort that under represented the populations the program now serves.

Mechanisms exist for a clinician to challenge an output, with nothing standing behind the output. LungAI's score is delivered inside the radiologist's existing viewer as an additional reference point that augments Lung-RADS, which remains the basis for documentation, and the radiologist reads and reports; LungIQ produces a ranked worklist that FQHC staff review before contacting patients. Those are review steps, and they are the whole of what is published: no accuracy commitment, warranty, indemnity or remediation position appears on the home, FQHC, imaging or business associate pages.

The exposure differs by component. A missed high risk patient on the LungIQ list is an omission nobody sees; a LungAI score on a borderline nodule can move a follow up decision in either direction, and the imaging page's own examples show scores that differ sharply from the Lung-RADS category. Once cleared, the company markets billing for the score, which will make the question of who stands behind it commercial as well as clinical.

Integration is deliberately light and one directional, with named partners on the imaging side. LungIQ starts from a flat file: FQHCs export demographics, notes, diagnoses and medications as CSV under a signed agreement, the company maps columns at configuration, and the FQHC page says no record system integration or IT project is needed. On the imaging side, LungAI's score appears inside the imaging center's existing CADe viewer through CADe partners, Coreline Soft among them, a PDF report with per nodule scores is written into the PACS, and GE, Philips and Siemens scanners are named as compatible.

That is read access by export and one directional writing into the PACS through a named partner. What holds it below the top is that nothing is written back into the FQHC's record, the flat file approach depends on periodic exports rather than live connection, and no marketplace listing documents a record system integration.

A hosted platform with no named host, region or residency commitment. The FQHC page shows the data path for identification: health centers export demographics, clinical notes, diagnoses and medications as flat files under a signed agreement, so full records leave the center for the company's environment. On the imaging side, scores are produced for display inside partner CADe viewers and a PDF report is written into the imaging center's PACS, which means scans or derived data reach wherever LungAI runs.

Neither path names a hosting provider, region or tenancy model on the home, FQHC, imaging or business associate pages, and the large language model that reads clinical notes may run with a third party.

The vendor's own price is reached only through a sales conversation. The home, FQHC, imaging and business associate pages publish no price, unit of charge or contracting model for LungIQ or LungAI. What they publish instead is the buyer's reimbursement: FQHCs can bill shared decision making visits under G0296 at their PPS rate, and imaging centers are told LungAI will be billable under CPT 0721T once cleared. The consortium model routes FQHC referrals to partner imaging centers, and who pays Oatmeal Health for that, and on what basis, is not stated.

Federally qualified health centers operate on tight margins, and nothing says whether the program costs them anything or whether its economics depend on a reimbursement that begins only after clearance.

Deliberately narrow coverage, named clearly, with the company's own validation behind part of it. One cancer, one screening modality in low dose CT, and three buyer types around it: federally qualified health centers that identify and refer eligible patients, imaging centers and hospitals that scan them, and pharma and medtech partners for trial work. Within that scope the whole pathway is covered, from eligibility through the scan to follow up.

Part of the coverage is validated. LungIQ is described as validated on a 35,000 patient FQHC panel, which is the setting it is sold into, and named partnerships span health centers and primary care associations in Oklahoma, Alaska, Massachusetts and Illinois. LungAI's validation is on National Lung Screening Trial data, whose population differs from the safety net patients the program serves, and the company states prospective performance has not been established.

The narrowness is a reasonable strategy: lung cancer screening has among the lowest uptake of any recommended cancer screening, concentrated where this company sells.