Abstractive Health

The summarisation model is the company. It grew out of two years of research with Weill Cornell Medicine on clinical summarisation, holds an issued patent for automated summarisation of a hospital stay using machine learning, and has taken NSF SBIR Phase I and Phase II awards for the summarisation tool specifically. Retrieval, question answering and alerting are all built on top of that engine rather than around a third party one.

This vendor does both things this index asks for and no one else in the category does either yet. It makes verification auditable at the finest granularity available, linking each sentence of a summary to the source note behind it, which is the Abridge Linked Evidence property applied to chart review rather than to note text.

And it publishes the rate, which is the Tortus property: it measured how often its own output raised a safety concern, put the figure at 6 percent of cases, named the failure modes as omissions and mischaracterisations, and concluded from its own data that a physician must stay in the loop.

Held short of a perfect score only by what is still missing: no confidence threshold, no routing rule and no abstention behaviour is published, so there is nothing that changes product behaviour when the retrieved record is thin. Note also that the free tier lets an individual clinician generate summaries with no institutional review, which is the free product governance inversion this index logged for Doximity.

The highest transparency in this category and the artifact is peer reviewed rather than promotional. The company published its summarisation pipeline and its full evaluation in JAMA Network Open, including the quantitative scores on which it was measured against physician written notes: ROUGE 0.322 versus 0.088, BERTScore 0.859 versus 0.796 and SCALE 0.691 versus 0.456.

It ships sentence level provenance, meaning every sentence of a summary links back to the source note that supports it, and it has published its own writing on why sentence matching matters in clinical summarisation. The product carries an in interface disclaimer that the assistant may occasionally make mistakes, and the company maintains a Trustworthy AI page. One gap remains and it is the standing index wide check: the underlying base model is not named on any retrieved surface. Ask whose model it runs on and whether patient content reaches a third party inference provider.

This vendor publishes the contract, which almost nothing else in the index does. A business associate agreement is published, alongside published network connection terms, a live public trust centre, a dedicated security page, an audited attestation and a public service status page.

A published agreement lets a counterparty read the obligations before entering a negotiation rather than discovering them in a redline, and a public status page is a small but real accountability artefact, since an outage or incident is visible to customers without the vendor choosing to disclose it. The surface is correspondingly large and one mechanism deserves naming because it is unusual.

This product retrieves records nationally rather than reading one organisation's chart, and geospatial retrieval around a patient's prior addresses reaches institutions with no relationship to the treating clinician, so a query is shaped by where someone used to live. That is a powerful way to assemble a history and it means the reach of a single lookup extends well beyond the care relationship that triggered it.

Held below the top grade because no explicit retention period, no statement on whether customer data trains or improves models, and no de identification posture was located, and for a product assembling a nationwide longitudinal record those three answers matter more than for a single site tool. Get them in writing.

The strongest evidence in this category and one of the better records anywhere in this index, because the headline result is unfavourable to the seller. Hartman et al, JAMA Network Open, 3 December 2024, doi 10.1001/jamanetworkopen.2024.48723, open access.

Cohort study of 1,600 emergency medicine records from 2023 admissions at NewYork Presbyterian and Weill Cornell Medical Center, funded by those institutions through the Joint Clinical Trials Office and Clinical and Translational Science Center rather than by the vendor. On conventional automated summarisation metrics the model was superior to physician written notes.

On a purpose built clinical framework it was marginally inferior on usefulness and safety, with safety concerns in 6 percent of cases arising from data omissions and mischaracterisations, and the authors concluded a physician in loop design is required. The chief executive disclosed his equity and his patent as competing interests inside the paper, the same candour this index credits in Lyssn.

Supporting evidence: NSF SBIR Phase I and Phase II, and a top three placement out of more than 150 entrants in the United States VA AI Tech Sprint, the same non commissionable external assessment that earned Knowtex credit. Honest caveats: single centre, cohort rather than randomised, the chief executive is first author, and the study covers one use case rather than the full commercial product. Buyers should also note the mismatch between the study's conclusion and the Clinically proven badge the company markets it under.

The compliance scaffolding is genuinely strong: SOC 2 Type II, a live public trust centre, a dedicated IT security page, a public service status page, a published business associate agreement and published Carequality connection terms.

The PHI surface is correspondingly large, because this product retrieves records nationally rather than reading one organisation's chart, and geospatial retrieval around a patient's prior addresses reaches institutions with no relationship to the treating clinician. Held at B rather than higher because no explicit retention period, no statement on whether customer data trains or improves models, and no de identification posture was located on any retrieved surface. For a product that assembles a nationwide longitudinal record, those three answers matter more than they do for a single site tool. Get them in writing.

A new top rung on this axis, above the previous benchmark. The existing ladder ran from HIPAA claimed with terms unpublished, up through a BAA explicitly included at no extra cost, up to coverage that is automatic and universal.

Abstractive Health does something none of those do: it publishes the business associate agreement itself as a linked public document, alongside its Carequality connection terms and separate United States and Canada terms of use. A buyer can read the actual instrument before creating an account rather than requesting it during procurement.

Publishing the text beats asserting the text exists, and this should be the benchmark cited on this axis going forward.

SOC 2 Type II claimed for the company itself, not for its data centres or its suppliers, which puts it on the right side of the supplier certification watchlist this index maintains. It specifies Type II explicitly, which answers the standing question this index asks of every SOC 2 claim, since Type I versus Type II is the whole assurance. It maintains a live public trust centre a buyer can open without contacting sales, a dedicated IT security page, and a public service status page.

Two precisions recorded without changing the grade: the company's own copy calls this a SOC 2 certification when SOC 2 is an attestation rather than a certification, and no ISO 27001 was located, which is the one artifact separating this from the fullest security package in the index.

A second pass located no clearance, device authorisation, pathway statement or exemption analysis. The flag raised in the earlier assessment is confirmed and sharpened, because the vendor has named itself into the category the question turns on.

Alerting is not an adjacent feature here. It is a named product line, listed alongside record retrieval, summaries and question answering, and the vendor uses the term clinical decision support for it. That is the exact category where United States law provides a conditional exemption, and one condition is that the professional can independently review the basis for the recommendation rather than relying primarily on it.

Two things make that condition harder to satisfy on this record than on a typical alerting product, and both come from the vendor's own account.

The first is architecture. Its founder has explained that feeding an entire medical record into a model produces inaccurate summaries, so the product summarises components separately and combines them. That is a sound engineering answer and it means an alert derived from the combined summary sits two transformations away from the source document. Establish whether a clinician can trace an alert back to the underlying record entry rather than to the summary.

The second is distribution. The product is sold self serve at a published per user price with a free trial, a free tier and a stated three minute setup. So a clinician can put a clinical alerting tool into practice with no institutional review, no governance committee and no validation against their own population. Nothing is being circumvented, and the checkpoint simply does not exist in the path.

Ask which framework the company believes governs the alerting module, and to see the analysis written down.

Publishing a peer reviewed study whose headline finding is that your own output is marginally inferior to physicians on usefulness and safety is the strongest governance signal available to a vendor, and naming the failure modes as omissions and mischaracterisations rather than reporting an aggregate accuracy number is exactly the disclosure this index has been asking the category for. The company also carries B Corporation certification and maintains a Trustworthy AI page.

Held at B for two reasons. No fairness, subgroup or demographic performance disclosure of any kind was located, which matters for a product assembling records across institutions serving very different populations. And the company describes billing nudges intended to optimise revenue under both value based and fee for service arrangements, which places it on the coding gradient this index tracks, although at the mild end and well short of the dedicated revenue modules graded down elsewhere.

The evaluation is peer reviewed, quantitative, and measured against the right comparator. The company published its summarisation pipeline and its full evaluation in a major medical journal, reporting scores on three named automatic metrics against physician written notes rather than against an internal baseline, and the physician note is the correct standard because it is what the summary replaces.

Publishing all three metrics rather than the most flattering one is itself a choice, since they measure different things and a vendor selecting among them controls the story. Sentence level provenance sits alongside: every sentence of a summary links back to the source note that supports it, which is the finest granularity of traceability found anywhere in this backfill and the level at which a fabrication actually becomes visible, because a sentence with nothing behind it has nowhere to point.

The company has also published its own writing on why sentence matching matters in clinical summarisation, and the product carries an in interface notice that the assistant may occasionally make mistakes, which puts the limitation where the clinician reads it rather than in a policy. Held below the top grade because no warranty, indemnity or remediation commitment attaches, and because automatic metrics measure similarity to a reference rather than clinical correctness. Ask for the omission rate on clinically significant content, and whose base model performs the summarisation.

The deepest interoperability record in this category and the reason is architectural rather than incremental. Most products in this index read the chart in front of them. This one is a direct implementer on the Carequality and CommonWell national networks and assembles a longitudinal record from outside institutions, performing geospatial retrieval around a patient's current and prior addresses.

It ingests CCDAs, FHIR R4 resources, PDFs and image based documents including handwritten scanned notes, normalises them into one structure, and leaves the original CCD and FHIR files downloadable rather than trapping them. It runs as a SMART on FHIR app, a standalone web app or a Chrome extension, exports as PDF, FHIR and JSON, and names twelve supported EHRs rather than claiming seamless integration in the abstract: AdvancedMD, athenahealth, Canvas, eClinicalWorks, Cerner, DrChrono, Epic, NextGen, Oscar, Greenway, Practice Fusion and Veradigm. One honest operational note the vendor discloses itself: provisioning access to the national networks takes about 72 hours.

Hosted software only. A web application, a SMART on FHIR embed and an API, with no on premise or customer hosted option described. The company publishes separate United States and Canada terms of use, which establishes operation in both jurisdictions, but no data residency commitment, hosting region, cloud provider or customer choice over where retrieved records are processed was located. For a product that pulls a patient's records from institutions across the country and holds them in one place, this is the significant unanswered question on the record.

A full published ladder with real figures, which is rare anywhere in this index and almost unheard of in products sold to health systems. A free forever tier with its limits stated plainly at five patient searches and summaries and five document uploads per month with unlimited question answering; a Standalone tier at 99 dollars per user per month; and a custom Organization tier for enterprise integration. A 15 day free trial of the full standalone product.

The company states explicitly that there are no setup or implementation fees, gives the setup time, and discloses the 72 hour wait for national record access. Its pricing FAQ names the catches rather than hiding them: staff need separate subscriptions, and the 99 dollar tier is capped at five total users so anything larger must move to enterprise contracting. That combination of a real price, a stated exclusion and a disclosed limit is the standard this axis should be measured against.

Genuinely broad by care setting rather than by specialty. Outpatient, inpatient and emergency care are all named, and the output types span SOAP notes, progress notes, transition of care notes, emergency department provider notes, discharge summaries and emergency to inpatient handoff notes, which is a wider set of documents than most of this category attempts.

Named customers reach across gastroenterology, pulmonary medicine, primary care, hospice and palliative care, and veterans community care. Graded B rather than A because the breadth is delivered by a general summarisation engine rather than by specialty specific behaviour, and no instrument level depth of the kind this index credits elsewhere was located.