Abstractive Health

The summarization model is the engine of the core module, and the platform sells record retrieval that stands without it. Abstractive retrieves a patient's records from the national exchange networks (Carequality and CommonWell), normalizes them and offers the original files for download; the company reports finding 94 percent of US patients with a 4 minute average retrieval, and describes its search as geospatial logic around current and prior addresses.

On top of that sit the learned parts: the summarization engine (in the company's published study, RoBERTa for content selection and Llama-2 7B for summarization, both fine-tuned), retrieval-augmented question answering, and chart alerts that clinicians define in plain language. Remove the model and retrieval is still a product.

The oversight structure is described and one part is missing. Every summary sentence and every answer links back to its source note, so a clinician can check each line; the question answering screen says the assistant may make mistakes; the terms require the customer to review all outputs; and the company says clinicians score each new feature or model change before launch and live outputs are monitored against a threshold that triggers investigation.

What is not published is what the product does when the retrieved record is thin or a sentence has weak support: no confidence display, abstention rule or escalation path. The free tier lets an individual clinician use summaries and alerts with no institutional review.

The approach and the research models are named, without version discipline for the product. The company's study names RoBERTa for content selection and Llama-2 7B for summarization, fine-tuned on 1,500 emergency records with k-fold cross validation, and its summaries page says the engine was trained on more than 10,000 patient charts with Weill Cornell Medicine and holds an issued patent.

The pipeline is described: section-to-meta summarization with sentence-level provenance, and retrieval-augmented question answering with citations. The company says every model change is evaluated by clinicians before launch, but the model versions in production and any change notice to customers are not published.

Substantial partial disclosure, with a short model chain. The company's published study used open models fine-tuned in house (RoBERTa and Llama-2 7B); hosting is on AWS; Guardrails AI is named as a validation partner; the terms name Carequality for exchange and Stripe for payments; and the business associate agreement binds subcontractors. There is no subprocessor list, and nothing says which model serves question answering and alerts today or whether any inference provider outside AWS receives patient content.

Published research short of independent validation, and a headline that runs past it. The home page says "Clinically proven" and cites the company's JAMA Network Open study (Hartman 2024): a retrospective single-site cohort of 1,600 emergency records at NewYork-Presbyterian and Weill Cornell, 1,500 used for training and 100 for testing, with a research model (RoBERTa and Llama-2 7B, fine-tuned on premise).

Automated scores favored the model (ROUGE 0.322 against 0.088), and on 50 summaries rated by three emergency physicians it was marginally inferior: usefulness 4.04 against 4.36 and safety 4.06 against 4.50, with no critical safety risk. The institutions funded the study, and the chief executive disclosed equity and a patent.

A second company study (NSF SBIR Phase I: two physicians, 12 patients, 60 true or false questions) reports correct answers rising from 60.0 to 83.3 percent with summaries, the source of the home page's 39 percent claim. The company also reports about 30 percent less pre-visit preparation time and a top three placement in the VA AI Tech Sprint.

Retention and de-identification are stated specifically, and the privacy policy reserves de-identified data for model work. The terms say the company deletes customer information within 36 hours of termination, and the published business associate agreement requires return or destruction of PHI within 30 days of termination and permits use of data de-identified under the HIPAA safe harbor.

The summaries page says the company does not use PHI or identifiable patient data to train its models, while the privacy policy says de-identified data is used for research and development of new products and to refine its algorithms and machine learning. No retention period during the contract is stated, and records are retrieved nationally from institutions near a patient's current and prior addresses, well beyond the treating organization. Beyond the Guardrails AI partnership and pre-launch evaluation, no safety engineering is published.

The business associate agreement is published and binds subcontractors. It is Appendix 3 of the terms of service (last revised 31 March 2026), readable before signup, and every account that accepts the terms accepts it, so it applies at every tier including the free one.

It requires subcontractors that receive PHI to agree in writing to the same restrictions, sets breach notice duties and return or destruction within 30 days of termination, and permits de-identified use under the safe harbor. The Carequality connection terms are published alongside. Which subcontractors those are is not listed.

SOC 2 Type II is named, and the report is released on request. The IT security page says the company maintains a SOC 2 Type II attestation and shares the report upon request as part of a security review, runs third-party penetration testing at least annually, monitors with Vanta, encrypts data at rest and in transit (TLS 1.3 where supported), and hosts on AWS.

The Vanta trust center is a JavaScript page that returns no content to a plain fetch, and the company's own pages give no report date or scope. A report released on the company's approval is a gated artifact.

No device claim is made, and the terms scope the product as support rather than advice. FDA's 510(k) database lists no clearance. The terms say the company is not a health care provider and gives no medical advice, and that its suggestions supplement rather than replace clinical judgment.

The clinical decision support module lets clinicians define alerts in plain language (for example, weight above dry weight with no recent loop diuretic), which is a recommendation function, and the company states no regulatory position for it. The free tier lets an individual clinician switch on alerts with no institutional review.

A named evaluation process, no subgroup results. The Trustworthy AI page says clinicians score every new feature or model change before launch on readability, completeness, curation and correctness (hallucinations, gaps, faulty logic and bias), plus usefulness and ethicality, against published 1 to 5 rubrics; live outputs are monitored with automated faithfulness signals and physician review; and Guardrails AI is named as a validation partner.

The company's peer reviewed study describes its patient mix (52 percent female, mean age 59.9) but reports no performance by group. Records assembled from many institutions vary in completeness, and nothing published breaks evaluation results out by age, sex, language or site of care.

A published error rate with its method and denominator. The company's own study (Hartman 2024, JAMA Network Open, with the chief executive as first and corresponding author under the company's affiliation) compares its research model's emergency handoff notes with physician notes: three emergency physicians rated 50 summaries, finding an incorrectness rate of 9.6 percent against 2.0 percent for physicians, five hallucinations all rated mild or no safety risk, mean safety 4.06 against 4.50 out of 5, and no critical safety risk.

The company cites the study on its home page. Every summary sentence links to its source note. The terms say the company is not responsible for the accuracy of output and require customer review, and its one indemnity covers intellectual property claims.

Named systems with read access and standards support, and no write path documented. The company is a Carequality implementer, lists CommonWell on its home page, and retrieves CCDA, FHIR R4, PDF and scanned documents into one normalized record, with the original files downloadable and exports in PDF, FHIR and JSON.

It lists 13 EHRs (AdvancedMD, athenahealth, Canvas, CareCloud, eClinicalWorks, Cerner, DrChrono, Epic, NextGen, Oscar, Greenway, Practice Fusion and Veradigm) under "Contact us for EHR integrations", describes SMART on FHIR delivery with Practice Fusion and athenahealth as examples, and offers backend integration "if supported". No marketplace listing or write-back into a record is described, and no production site is tied to a named EHR.

A single hosted option with the location implied. The company hosts on AWS and publishes separate US and Canada terms; the product runs as a web app, a SMART on FHIR app, a Chrome extension and an API. No region, residency commitment or tenant model is stated, and records retrieved from institutions nationwide are stored in the company's cloud under the business associate agreement.

Published tiers with figures, a unit of charge and a route to start without a sales call. A free clinician account (unlimited question answering, five patient searches with summaries and five PDF uploads a month), Standalone at $99 per user per month for up to five users, and an Organization plan at custom pricing, required above five users, with API access and EHR integration. A 15 day free trial covers Standalone. The company states no setup or implementation fees for Standalone, a three minute setup, and about 72 hours to provision national record access. Organization pricing is not published.

Settings are named with validation behind part of them. The company names outpatient, inpatient and emergency use, pre-visit review, handoff notes, transitions of care and discharge summaries, and population work for the CMS ACCESS model; its case studies cover reproductive endocrinology and VA community care. The peer reviewed validation covers emergency to inpatient handoff notes at one site, and the SBIR study outpatient chart review by two physicians. No age, specialty or population exclusions are published.