Mednition

The model is the product and it does something the existing process cannot. Standard triage runs on a rules based severity index applied by a nurse under time pressure; this system reads the same intake data plus the free text of the nurse's own notes and compares it against patterns from millions of prior visits.

Reading unstructured triage notes is the part that makes this irreducibly a model. Those notes are written in seconds, heavily abbreviated and inconsistent between nurses, and the signal the company claims to extract from them, a patient who looks stable but is not, is precisely what a rules engine on vital signs misses.

Decision support with the boundary stated plainly and repeatedly: the system recommends an acuity level and supports the nurse's judgement rather than replacing it. The nurse assigns the acuity.

The design choice worth crediting is the deliberate absence of intrusion. The company states KATE adds no new screens and requires no workflow change, which is the correct response to the environment: a triage nurse cannot stop to consult a second application, and an alerting product that demands attention at that moment competes with the patient in front of them. Held at B because nothing published describes what happens when the recommendation and the nurse disagree, whether the divergence is recorded, or what a nurse is shown to justify a recommendation they are being asked to weigh in seconds.

Substantially better than the category norm because performance is published with denominators in preprints carrying named authors, not only in marketing copy. The later sepsis model is reported at 99 percent area under the curve with 95 percent sensitivity and 96 percent specificity, on a stated cohort of 540,884 patients containing 14,676 positive cases across 16 sites, against a named academic sepsis definition. An earlier model is separately reported at 94 percent.

Stating the positive case count alongside the cohort size is the disclosure that matters in a rare event problem, because a headline accuracy figure on a heavily imbalanced population can be meaningless without it.

Held below A because the work sits in preprints rather than peer reviewed journals, because no architecture or feature description is published, and because the results are hosted on the company's own research page rather than assessed by anyone outside it.

Nothing identifies any party in the chain: no model or model family, no architecture or feature description, no hosting arrangement and no sub processor list was located in two passes, and no position on whether customer data improves the models was found in either direction. One aspect of the input deserves naming because it is easy to overlook and it changes what the unnamed chain is handling.

This model reads free text triage notes, which are among the least guarded documents in a record. They are written fast, under pressure, usually unreviewed, and they routinely carry social detail, clinical suspicion and description a nurse would phrase differently in a formal note. A system consuming them at scale is consuming a category of text nobody drafted expecting an audience beyond the next clinician on shift.

That makes the training question sharper rather than routine, because a de identification process built for structured fields does not address what a triage note says about a person's circumstances, and it makes the identity of every party handling that text something a buyer should establish rather than infer. Ask for a sub processor list, the hosting arrangement, and an explicit position on whether triage note text is used to train or improve models.

The retrospective evidence is genuinely strong for its type: a multi site cohort of more than half a million patients with the positive case count stated, and comparative improvements against standard screening reported separately for sepsis, severe sepsis and septic shock rather than collapsed into one figure. Systemwide adoption at a named health system with four named sites corroborates it in practice.

Two things hold it at B. Everything located is retrospective, so nothing demonstrates that deploying the model changes what happens to patients rather than that it would have flagged them. And the operational claims that would show that, a doubling of high risk identification at triage and a one day reduction in length of stay, are asserted without a study, a baseline or a period attached.

Graded on an honest basis. No published retention schedule, encryption detail or position on model improvement from customer data was located in this pass.

One aspect of the data is worth naming because it is easy to overlook. This model reads free text triage notes, which are among the least guarded documents in the record: written fast, unreviewed, and often containing social detail, suspicion and description a nurse would phrase differently in a formal note. A system consuming them at scale is consuming a category of text nobody drafted expecting an audience.

Graded on an honest basis and flagged for re verification. No compliance statement or agreement posture was located in this pass.

Systemwide deployment at a large health system means agreements exist and have been negotiated at scale; they are simply not public. A buyer should ask directly rather than infer anything from the absence.

Recorded honestly: the dedicated trust and security search this index requires was not run in this pass, so the grade is provisional and should not be quoted until it has been. No attestation was encountered incidentally.

The company states it is funded in part by major health systems, which if accurate means investors with their own security standards have examined it, but that is an inference rather than a finding.

Breakthrough Device Designation only, and this index is firm that designation is not clearance. The programme expedites review of a device that has not yet been authorised; it confers no permission to market as a diagnostic and involves no finding that the device is safe or effective.

What makes this record instructive is the elapsed time. The designation was granted in November 2023 and no clearance was located as at August 2026, nearly three years later. A comparable record in this index shows the other outcome, where designations converted into three separate clearances. The useful question on this axis is therefore not whether a vendor holds a designation but whether its designations have ever converted, and here that answer is not yet.

The product is meanwhile marketed as decision support for a nurse, which is a position that may not require clearance at all, so a buyer should establish which claim the company is relying on.

The company has raised equity as an aim, describing its work as advancing equitable care, and it names the failure mode its category is known for rather than avoiding it: that high sensitivity in sepsis detection is easily bought at the cost of specificity, producing false positives and alert fatigue. Naming that tradeoff explicitly is more candour than most vendors offer.

What is absent is the data behind the equity claim. No subgroup performance was located, and the application makes that omission consequential. Emergency triage acuity assignment is one of the best documented sites of disparity in medicine, with differences recorded across race, language and presentation, and a model trained on historical triage decisions learns from exactly those assignments. A system could reduce that variation or entrench it, and the published material asserts the first without showing it.

This is the counterexample to the measurement failure recorded across the detection products in this index, and it is worth reading against them. Performance is published with denominators in preprints carrying named authors rather than in marketing copy.

The later sepsis model is reported at 99 percent area under the curve with 95 percent sensitivity and 96 percent specificity, on a stated cohort of 540,884 patients containing 14,676 positive cases across 16 sites, assessed against a named academic sepsis definition.

Stating the positive case count alongside the cohort size is the disclosure that matters most in a rare event problem, because a headline figure on a heavily imbalanced population is close to uninterpretable without it, and it is precisely the decomposition the monitoring vendors in this index decline to give. Naming the reference definition matters equally, since sepsis has several and a model looks better or worse depending on which is used.

A buyer can therefore argue with this claim, which is what makes it a claim. Held below the top grade for three reasons rather than one. The work sits in preprints rather than peer reviewed journals, so no external referee has examined the method. The results are hosted on the company's own research page rather than assessed by anyone outside it. And no warranty, indemnity or remediation commitment attaches, so the disclosure is transparency rather than accountability. Ask whether the deployed model is the one evaluated, and whether the published operating point is the one shipped.

Integration into the record is described as a defining property rather than a feature, with the company positioning the platform as record integrated and stating that it adds no new screens for the nurse. That is a strong claim operationally: the output has to appear inside the triage workflow the nurse is already in, within the seconds that workflow allows.

Systemwide deployment at a health system across multiple hospitals corroborates that the integration works at scale. Held at B because no named record vendor, interface standard or integration mechanism was located.

Not described. No hosting model, region, retention schedule or customer controlled option was located.

The latency requirement is the constraint that shapes any answer, since a recommendation that arrives after the nurse has assigned an acuity has no value. How that is achieved, and where triage text travels to be processed, is not published.

Nothing published. No price, no mechanism and no unit of sale, and no indication whether the sepsis model, the triage acuity model and the paediatric product are licensed together or separately.

The value case the company advances is unusually measurable, which makes the silence more noticeable rather than less. A stated one day reduction in length of stay and a doubling of high risk identification are both quantities a hospital already tracks and could price directly against a per visit or per site fee, if one were published.

Deliberately narrow and coherently so. The setting is emergency department triage and essentially nothing else, covering adults and, through a separate product, children, with the company framing the emergency department as the front door where most admissions originate.

That focus is the strength and the boundary. The product is specialty agnostic within emergency medicine because triage handles undifferentiated presentations, and it does not extend to the ward, the intensive care unit or ambulatory settings. Deployment appears to be United States only.