Layer Health

The language models are the engine of the core product, and a services line now sits beside them. The platform reads the full patient record, structured and unstructured, and suggests registry and quality measure answers with the supporting evidence from the chart; the clinical pathways and site of care products screen the record for procedure eligibility. The registries page also offers AI Transitionary Services, in which the company's own clinical experts use the platform to do a health system's abstraction on its behalf, moving to automated submission field by field as validation thresholds are met, and the company recruits contract abstractors for that work.

Remove the models and the company still sells outsourced abstraction by its own staff, with far less to show for it. The record's earlier statement that nothing sits underneath the models no longer matches the offer.

A strong oversight structure is described, and the company's own copy tells buyers to trust every output. The customer's abstractors validate the model on their own data before go live; answers arrive as suggestions with the chart evidence behind them; and fields move to automated submission only when validation reports show they meet a required threshold, with the company's own abstractors available to run the work in the meantime. That is a boundary published both ways, with the threshold values unpublished.

The registries page heads its accuracy figure with "Trust every data point." That is a universal attached to correctness, and it grades here whatever the validation design, because a quality team told to trust every value will not keep the review step that catches the fields the model gets wrong; the same page's 98 percent validated accuracy says some are wrong. For fields that move to automated submission, that review step is the only one left.

The approach is named in general terms and nothing is identified. The platform applies large language models to the full record, is said to need no labeled training data, returns each answer with its chart evidence, is continuously monitored and improves with every piece of feedback; the home page says it is built on the founding team's more than 200 peer reviewed publications, from MIT, Harvard Medical School and Google backgrounds.

What is missing is which models, what they were trained or tuned on, and a version practice that tells a customer when a model changed. Continuous improvement from feedback makes that last question sharper, since a field validated before go live may be served by a different model later. The team's publications are its research record, not an evaluation of this platform, and no architecture or evaluation method for the product is published.

The architecture is described and no party in the chain is named. The platform applies large language models to the full record, connects to Epic and Cerner through a FHIR first approach with no custom development, and returns each answer with the chart evidence behind it. Which models it uses, whether built, fine tuned or licensed, where they are hosted, and which subprocessors touch patient records are not stated on the home page, registries page, privacy policy or terms; the privacy policy covers website visitors and names only analytics providers.

The company says the platform needs no labeled training data, which is a claim about human annotation, not about what the models were trained or tuned on; its chief executive has described the models as trained on longitudinal patient data, and the home page says the AI improves with every piece of feedback. Whether customer chart data or abstractor feedback flows into a model other customers use is the party question here.

Named deployments with outcome figures and an outline of how they were measured. The company reports 50 to 60 percent faster registry abstraction, 98 percent or better validated accuracy on each health system's data before go live, and go live within ten weeks; Froedtert and the Medical College of Wisconsin reported abstraction time cut by more than 65 percent in the March 2025 Series A announcement, and White Plains Hospital is the subject of a published case study on halving abstraction time. Intermountain Health (33 hospitals, stroke, bariatric and cardiovascular registries), Johns Hopkins Medicine and MultiCare are named customers. The registries page describes the method: abstractor time per case is baselined in the current workflow and ground truth established before the AI is switched on.

No denominator, field count or registry split accompanies the accuracy figure, the time savings carry no stated period, and no peer reviewed evaluation of the platform is published. The home page's claim that the models perform just as well as humans or better has no published comparison protocol.

General assurances and one structural control, with no commitment on what the models learn from. The company states SOC 2 Type II certification and compliance with United States health privacy law in its announcements, the footer says enterprise grade security and compliance, and every abstracted value carries a pointer to the note it came from, so a fabricated value has nowhere to point. None of that is a commitment about data use.

No retention period, deidentification step or training use statement is published on the home page, registries page, privacy policy or terms; the privacy policy covers website visitors, and lets the company deidentify and share their information. The home page says the AI improves with every piece of feedback, and the chief executive has described the models as trained on longitudinal patient data, so whether one health system's charts or abstractor corrections improve models another uses, and how long charts read in bulk persist, are unanswered. No safety engineering is published beyond evidence linking and pre go live validation.

Compliance with United States health privacy law is asserted in company announcements alongside the security certification. No body certifies HIPAA compliance, so that assertion reports intention rather than an audited finding, and it is graded accordingly.

The company publishes no business associate agreement posture, notice or compliance documentation. The customer base of large academic and regional health systems implies these terms are negotiated and in place, since no organization of that size deploys across 33 hospitals without them, but the public record does not show it.

SOC 2 Type II certification is stated in company announcements, with the report type named rather than left as an unqualified claim.

That is the whole of it. No trust center or security portal, penetration testing statement, ISO 27001 or HITRUST certification appears on the home page, registries page, privacy policy or terms, and nothing says whether the report can be requested by a prospective buyer; the site's footer says only enterprise grade security and compliance. For a company selling into large health systems whose vendor risk processes will demand the report, making its availability visible would be a cheap improvement.

No device claim is made, registry abstraction is a data operation, and two newer products reach toward clinical decisions with no position stated. Abstracting a chart into a registry or quality field reports what is documented. The clinical pathways product is described as reasoning across the full record to identify eligible patients for high value procedures and turning data into decisions for guideline aligned care, and site of care optimization screens surgical cases against protocols; a surgeon quoted on the home page says the tool found a patient scheduled for an ambulatory surgery center who was ineligible and is already prompting real clinical decisions. Each answer comes with its chart evidence for a clinician to review, so this is support with a reviewable basis, not a diagnosis, and no regulatory position for either product appears on the home page, registries page, privacy policy or terms.

Registry work answers to registry definitions and accreditation bodies; the pathways and site of care products are where a stated position would matter.

An evaluation process written for each customer is described, and no subgroup result is published. The registries page says every deployment is validated on the health system's own data before go live, that customers track model validation reports until each field meets its required threshold for automated submission, and that performance is baselined against abstractors' own time and ground truth; the home page adds that models are continuously monitored for sustained performance. Validation reports produced for a customer's own quality team to accept or reject a field are material written for the customer's review.

Results are not compared across sites, specialties, note styles or patient groups. Registry data does not stay in the registry; it feeds national quality measurement, public rankings, reimbursement adjustments and research, so if abstraction accuracy varies with documentation style, note length, specialty or language, the error is written into a permanent record and inherited silently by everything computed from it. No cross site or subgroup result is published, though the per customer validations are exactly the studies that would show it.

Mechanisms exist to check each answer, and nothing stands behind it. The home and registries pages say every answer is sourced from the record with the supporting evidence from the chart, that suggested answers are validated on each health system's data before go live, and that fields move to automated submission only once validation reports show they meet a required threshold. That is how an answer can be challenged: an abstractor can see where each value came from.

The company publishes 98 percent or better validated accuracy without a denominator, field count or registry split, so it is not an error rate a buyer can test. No warranty, indemnity or remediation commitment appears on the home page, registries page, privacy policy or terms, which cover the website only and cap liability at $100. A wrong value submitted to a national registry carries into quality rankings and research with no published route back to the vendor.

Named record systems through a standard, with write back undescribed. The registries page says the platform supports Epic and Cerner through a standardized, secure FHIR first approach, connected to existing infrastructure with no custom development, and works inside the abstractors' existing workflow; running across 33 Intermountain hospitals and at several other health systems and a national research organization requires real access to charts at scale rather than document upload. Delivery is also described as application programming interface access or modules integrated with the record.

No marketplace listing or integration documentation is published, what the platform writes back into the record (if anything) is not described, and registry submission involves a second set of interfaces to the registry bodies that is not described at all.

Nothing is published about where the platform runs or where charts rest. The registries page describes connection to Epic and Cerner through a FHIR first approach with no custom development, and the footer says enterprise grade security and compliance; no hosting model, cloud provider, region, on premise option or retention schedule appears on the home page, registries page, privacy policy or terms. The privacy policy's statement that personal information may be transferred and accessed from around the world covers website visitors, not the charts the platform reads.

Where the data sits matters a good deal, because the workload is bulk historical chart reading across many hospitals: Intermountain alone runs it across 33, and a national research organization uses it on patient data drawn from institutions across the country. Whatever the architecture, a large volume of longitudinal records is processed somewhere, and the public record does not say where or for how long it persists.

No price is published, and how the product is sold is discoverable. The registries page offers three models: AI native software used by the health system's own abstractors, AI transitionary services in which the company's clinical experts do the abstraction on the platform until fields qualify for automated submission, or a mix by registry; implementation runs from contracting to go live in ten weeks or less. Other products (custom quality measures, clinical pathways, site of care optimization) are sold to health systems and research organizations, and the route in is a demo.

What a buyer cannot establish before contact is the unit: per chart abstracted, per registry, per hospital or per seat, which scale very differently for a system running several registries across many hospitals. The comparison a buyer needs is against the current cost of nurse abstractors, which every hospital already knows.

Broad within chart abstraction, with validation behind part of it. The buyer is a health system's quality, registry or clinical data team, and the platform covers registries across cardiovascular, surgery (including NSQIP), oncology, stroke and heart failure, state cancer registries and custom registries, plus custom quality measures such as surgical site infection; newer products screen records for procedure eligibility (clinical pathways) and surgical site of care. Research abstraction is a second setting, with a national cancer organization using the platform on patient data drawn from across the country.

Results are published for registry abstraction at named health systems (Froedtert, White Plains Hospital, Intermountain), while the pathways and site of care products carry revenue figures and a surgeon's testimonial rather than validated results, and each new registry brings its own definitions and field rules, so coverage grows registry by registry. Geography is the United States.