Prenosis

The authorised device is an algorithm and nothing else. There is no monitor, no scanner and no workflow platform, and the thing a hospital buys is a score computed from inputs it already generates.

The biobank complicates the reading and, on inspection, supports it. More than 100,000 blood samples from over 25,000 patients, gathered across ten hospitals over a decade, is a formidable proprietary asset, and it is training data rather than product. What is sold is the model that dataset produced. The company's stated ambition, to move from a risk score toward recommending personalised therapy in real time, is a model ambition rather than a data brokerage one.

A published design decision here is worth more than most vendors' oversight sections, and it is stated in one line: it is not an alert system.

That is a deliberate refusal rather than an omission. Sepsis prediction has an unusually well documented failure history in which alerting systems fired constantly, clinicians stopped reading them, and the tool became noise. This product instead returns a score and a placement in one of four risk categories, leaving the clinician to interpret it alongside the patient in front of them. A vendor declining the more autonomous design in a category where alerting is the norm, and saying so plainly in its own regulatory announcement, is making a claim it can be held to.

Held at B because what a clinician is instructed to do at each of the four risk levels is not published, and because the score's relationship to existing sepsis protocols and bundle timing is not described.

The De Novo route forces more disclosure than a clearance does, because creating a new device classification requires the agency to publish the special controls that will govern every subsequent device of that type, and that document is public.

What is disclosed in company material is the input count and composition, up to 22 parameters spanning measured biological markers and clinical data, and the output structure of a continuous score plus four discrete risk categories tied to named outcomes. Development and validation have been published in the peer reviewed literature.

Held below A because the model class itself is not named in public material. A directly comparable record in this index publishes its algorithm family and its default alerting thresholds outright, so the bar has been set within the category and this vendor sits just under it.

Nothing identifies any party in the chain: no model or model family, no hosting arrangement and no sub processor list was located in two passes, and no published position on customer data handling, retention or model improvement was found. The provenance question on this record is sharper than anywhere else in this index, because what underpins the model is not data but tissue.

The company's biobank holds more than one hundred thousand blood samples from over twenty five thousand patients in a physical laboratory, alongside their clinical records. That differs from a data corpus in three ways that matter. Blood is biological material, so holding it is a custody question rather than a storage one.

It remains identifiable through what can later be measured from it, including things nobody has thought to measure yet, so de identification does not settle the question the way it might for text. And consent to donate a sample for indefinite retention and future research is a different instrument from consent to have a chart used, obtained at a different moment and usually from a patient who is acutely unwell. Nothing published describes what those patients agreed to, how long samples are held, whether they can be withdrawn, or what happens to them if the company is sold. Ask all four.

The strongest evidence position in this category and one of the strongest in the index, built from four independent kinds of validation.

De Novo authorisation is the most demanding of the routes available to a novel device, since there is no predicate to lean on and the agency must be persuaded the type of device is safe and effective at all. Development and validation have been published in the peer reviewed literature. A separate peer reviewed comparison reports the score predicting sepsis, mortality and deterioration better than clinical scores and widely available biomarkers, which is the comparison that matters and one most vendors avoid. A federal agency has committed 20 million dollars to fund a randomised controlled trial of 800 patients, which is a prospective trial paid for by a body with no commercial interest in the result.

There is a fifth and unusual form: a competitor subsequently obtained clearance using this device as its predicate, meaning the regulator treated this evidence base as the reference standard for the category.

Graded on an honest basis. No published stewardship position covering customer data handling, retention or model improvement was located in this pass.

The provenance question is sharper here than anywhere else in this index because what was collected is not data but tissue. The biobank holds more than 100,000 blood samples from over 25,000 patients in a physical laboratory, alongside their clinical records. Blood is biological material, it remains identifiable through what can later be measured from it, and consent for its collection and indefinite retention is a different instrument from consent to use a chart. Nothing published describes what those patients agreed to, how long samples are held, or whether they can be withdrawn.

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

The structure is more complicated than a software vendor's and worth establishing before contracting. This company operates a physical laboratory holding patient specimens, integrates with hospital records, and distributes commercially through a large diagnostics manufacturer's platform, so a hospital's data may touch three parties. Which of them holds which obligation is not described publicly.

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 trust centre or attestation was encountered incidentally.

One consideration that will bear on the eventual grade: distribution through a major diagnostics manufacturer's algorithm platform means this product has passed that company's own supplier assessment, and a manufacturer of that size does not add third party algorithms to a clinical platform without one. That is real external scrutiny sitting outside the usual attestation framework.

The strongest regulatory position of its type available. This was the first artificial intelligence sepsis diagnostic ever granted marketing authorisation, obtained through De Novo on 3 April 2024, and the company is explicit that the route was necessary precisely because no authorised predicate existed.

The consequence is the part worth recording. A De Novo does not merely authorise one product, it creates a classification with special controls that subsequent devices of the same type can clear against. A competitor later obtained a 510(k) using this device as its predicate, so this authorisation is now the regulatory foundation of the cleared sepsis category rather than one entry within it.

The device is also described as combining diagnostic and predictive claims, which the company states had not previously been available together in a legally marketed sepsis device.

No subgroup performance or monitoring policy was located in company material, and the grade is C rather than lower because the underlying documentation very likely exists: De Novo special controls and a peer reviewed development and validation paper both ordinarily carry demographic performance, and neither was retrieved in this pass. Re verify against the classification order and the publication.

The context makes this worth pursuing rather than assuming. Sepsis prediction is the area of clinical artificial intelligence with the most publicly documented validation failure, after a widely deployed proprietary model was found in independent evaluation to perform far worse in practice than its developer claimed. Any vendor in this category is working in the shadow of that episode, and published subgroup performance is the specific thing that would distinguish it.

The regulatory route here produces more public disclosure than the usual one, and the distinction is worth understanding because it applies to every product authorised this way. Creating a new device classification requires the agency to publish the special controls that will govern every subsequent device of that type, so the authorisation generates a public document describing what such a device must do to be acceptable, rather than only a summary of one submission.

A buyer therefore has an externally set standard to read alongside the vendor's own material, and a competitor entering later must meet it. Company material adds the input count and composition, up to twenty two parameters spanning measured biological markers and clinical data, and the output structure of a continuous score plus four discrete risk categories tied to named outcomes, so a clinician knows what feeds the score and what the categories mean.

Development and validation are published in the peer reviewed literature. Held below the top grade because the model class itself is not named in public material and no warranty, indemnity or remediation commitment attaches. A directly comparable record in this index publishes its algorithm family and its default alerting thresholds outright, so the bar has been set within the category and this vendor sits just under it. Ask for the model family, the default thresholds, and calibration in the population you serve.

Integration into the record is described as a property of the product rather than an option, with the score surfacing where clinicians already work, and a named health system deployment confirms it operates in live care.

The distribution route adds a second and unusual integration path. The product is offered through a major diagnostics manufacturer's algorithm platform, which already connects to hospital laboratory and record infrastructure, so a hospital may reach this model through a channel it has already integrated rather than through a new connection. That is a meaningfully different adoption route from every other deterioration vendor here. Held at B because no interface standard or named record vendor certification was located.

Not described. No hosting model, region or retention schedule for computed scores was located.

The architecture has a physical dependency the other records in this category do not: the score requires biological markers measured from a blood sample, so it cannot run on record data alone and its availability is tied to laboratory turnaround. That is a real operational constraint in an emergency department, where the value of an early warning depends on how early it arrives, and nothing published states the expected time from draw to score.

Nothing is published: no price, no mechanism, and no indication whether the model is per test, per patient or per facility.

One structural reason is worth recording rather than treating the silence as simple reticence. The product is distributed commercially through a large diagnostics manufacturer's platform, so the price a hospital pays may be set inside that manufacturer's contracting rather than by this company, and the vendor may genuinely not be the party able to publish it. A buyer should establish which route they are purchasing through, because the answer determines who they are negotiating with. Worth asking too whether the biological markers used are already run in the hospital's own laboratory, since if so the marginal cost is the score rather than the testing.

Deliberately narrow. The population is adult acute care patients with suspected serious infection, reached in the emergency department and on inpatient wards, and the product does not extend to ambulatory care, paediatrics or any condition outside its indication.

That narrowness is a design position rather than a limitation, since sepsis is the single largest driver of in hospital mortality and the tool is built to do one thing at the moment it matters. A federally funded trial in severe respiratory infection signals where the indication may widen next, and until that reads out the reach is what the authorisation says it is.