Bayesian Health

The model is the product and the company is explicit about the architectural claim behind it. Real time machine learning models run inside the EHR, learning from messy multimodal structured and unstructured data streams, and the company positions its approach as adaptive rather than one size fits all, taking into account patient population diversity, how clinicians actually deliver care at a given site, and the characteristics of each health system. That per site adaptation is a substantive design position, not a marketing line, and it is what the published research attributes the accuracy and adoption gains to.

The system alerts rather than acts, which is the correct and conservative posture for deterioration detection, and the company treated clinician acceptance as a research question rather than an assumption. One of its three published studies specifically examined clinicians' experiences with the deployed system and concluded that human machine teaming is key to adoption, with over 4,000 caregivers participating in the prospective deployment. Studying whether providers trust and act on the alert, and publishing the result, is a materially higher standard than shipping an alert and reporting an AUC.

Transparency here comes through the peer reviewed literature rather than a datasheet, which is the stronger form. Three studies published in Nature Medicine and npj Digital Medicine describe the approach, the deployment, and the clinician experience, with cohort sizes and settings disclosed, so an external reviewer can examine the methodology directly.

The company also publishes the specific claim that its approach produces fewer false alerts than prior systems, which is the metric alert fatigue turns on and one that competitors rarely commit to in print.

Nothing identifies any party in the chain: no model or model family, no foundation model provider, no hosting arrangement and no sub processor list was located in two passes, and no published stewardship framework or data governance terms were found on the company's own materials.

The access surface is among the broadest in this index, since the platform reads live records including clinical notes across every hospital department, so whatever unnamed parties are involved are handling the whole record rather than a slice of it. The clearance carries a partial and easily overstated answer that should be read precisely. A device submission necessarily addresses what data the device consumes and how it behaves, so some of this exists in a regulator's file.

It says nothing about what the company does with that data outside the device's operation, which is exactly what this axis asks. The per institution adaptation makes the central question unavoidable: a model tuned to one hospital's population is built from that hospital's records, and whether anything derived from it reaches another customer is a term to settle in the agreement rather than infer from architecture.

Ask that, ask what is retained and for how long, who inside the company can reach identified records during implementation or incident investigation, and what is returned or destroyed at contract end.

The strongest clinical evidence base of any vendor in this index. Three large prospective multi site cohort studies conducted with Johns Hopkins across five academic and community hospitals over a five year period, spanning 764,707 patient encounters of which 17,538 involved sepsis, with 2,000 plus providers using the software during 2.5 years of prospective deployment.

Reported outcomes include roughly 20 percent lower likelihood of sepsis death, detection hours earlier than traditional vital sign scoring, and one in three cases caught by the system first. The company's own framing is notable and defensible: this was presented as the first time lives saved were associated with a clinically deployed AI platform. Buyers should still note the studies were conducted in collaboration with the founder's own academic institution.

Converted from Not Rated after a second search. No published stewardship framework or data governance terms were located on the company's own materials, and the prior note's characterisation of the data surface stands: the platform reads live records including clinical notes across every hospital department, which is among the broadest access surfaces in this index.

One indirect signal is now on the record and should be weighed without being over credited. The sepsis module went through device clearance, and the company describes years of work with the regulator validating performance across diverse hospital settings and patient populations and assessing missed cases and alert fatigue, together with a post market monitoring and quality assurance programme. A submission of that kind necessarily addresses what data the device consumes and how it behaves. It does not address what the company does with that data outside the device's operation, which is what this axis asks.

The unanswered questions are ordinary and important. Whether data from a health system is used to develop or improve models served to other health systems. What is retained, in what form, and for how long. Who inside the company can access identified patient records during implementation, support or incident investigation. And what is returned or destroyed when a contract ends.

The per institution adaptation the company describes makes the first of those unavoidable. A model tuned to one hospital's population is built from that hospital's records, and whether anything derived from it reaches another customer is a question a health system should settle in the agreement rather than infer from architecture.

Converted from Not Rated after a second search. No statement of business associate status, agreement availability, terms or subprocessors was located. The prior note's reasoning holds: live record integration at named health systems means business associate arrangements are structurally required and certainly exist contractually, and the gap is that nothing is published.

The device clearance obtained in April 2026 does not change this and should not be read as though it does, because a buyer may reasonably assume that regulatory review covers everything. It does not. Device regulation examines whether a product is safe and effective for its intended use, including its cybersecurity design. The health privacy rule governs a separate question, which is the terms on which a vendor may receive, use and disclose patient information on a covered entity's behalf. A company can hold a clearance and have an inadequate business associate arrangement, or the reverse. The two regimes do not substitute for one another and a hospital's procurement process should treat them separately.

What to establish, and it is straightforward for a vendor already operating at named health systems. Which legal entity signs, what the agreement permits by way of use for product improvement, whether subprocessors are named and how changes to that list are notified, what the breach notification commitment is in practice, and what happens to the institution's data and to any institution specific model artefacts at termination.

That last item connects to this vendor's own architecture rather than being generic, because the company describes adapting models per health system, and an adapted model is a derivative of that institution's records.

Converted from Not Rated. No SOC 2, HITRUST or ISO 27001 attestation was located and there is no trust centre, and the prior note was right that this is a conspicuous asymmetry for a company with so much published clinical rigour. Something has since changed that this axis should recognise.

The sepsis module received device clearance in April 2026. Since 2023, a submission for a device that connects to a network or contains software of this kind must include a cybersecurity plan, a software bill of materials, and a description of processes for monitoring and delivering post market security updates and patches. A device cleared in 2026 has therefore had its security documentation examined by a federal regulator as a condition of reaching the market.

That is a genuine form of security assurance and it differs in kind from a commercial attestation. An attestation examines an organisation's control environment against a framework and is available to a counterparty on request. A device submission examines a specific product's security design and lifecycle plan, is reviewed by a regulator with enforcement powers, and is not published. Neither substitutes for the other, and most vendors in this index have neither.

Held at B rather than higher for three reasons. It covers the cleared module, not the wider real time clinical intelligence platform that reads live records across every department. The documentation is not public, so a buyer cannot read it and must ask. And there is still no organisational attestation covering the company's own environment, development practices or staff access.

Ask for the cybersecurity documentation submitted with the clearance and the software bill of materials, the post market update process, and separately whether any organisational attestation exists for the platform beyond the cleared module.

Upgraded from Not Rated, and the prior note is superseded rather than refined. It recorded that no clearance was claimed and none was located, and identified this as the axis to examine most carefully because the product sits close to the device boundary. That judgement was right and the boundary has since been crossed deliberately.

On 30 April 2026 the sepsis flagging module received 510(k) clearance, announced in May. The clearance is specific and the details matter. It is classified under the regulation covering software devices to aid in the prediction or diagnosis of sepsis, is prescription use only, and was found substantially equivalent to an earlier de novo authorised sepsis product. Both are described in the submission as adjunctive tools rather than standalone diagnostic solutions. Intended use covers risk assessment in the emergency department, intensive care unit and ward, evaluated within the twenty four hours surrounding sepsis onset across several thousand encounters at four hospitals, with unit, integration and system level testing and human factors validation documented. The company had previously held breakthrough device designation, and the submission was anchored on a prospective study published in a leading clinical journal.

Why this earns the top grade rather than merely a good one. Most sepsis prediction software in United States hospitals operates as unregulated decision support, on the reasoning that a clinician can review the basis independently. This company took the opposite path, spent years with the regulator establishing sepsis definitions, validating performance across settings and patient populations, and assessing missed cases, alert fatigue and post market monitoring. A buyer can read the clearance record rather than accepting a regulatory rationale from the vendor.

Two things to hold in view. The clearance covers the sepsis module, not the wider clinical intelligence platform, so confirm which components are cleared. And a reimbursement pathway decision was pending at the time of review.

Better engagement with this than almost any vendor in the index, though it is embedded in the technical approach rather than presented as a governance program. The company's core stated argument is that one size fits all models fail because they ignore patient population diversity and site specific care delivery patterns, and its adaptive approach is the response to exactly that failure mode.

The prospective studies span both academic and community hospital settings, which tests generalization across meaningfully different populations. Falls short of an A because no formal governance framework, monitoring commitment, or subgroup performance breakdown was located.

This is the strongest combination on this axis located anywhere in the backfill so far, because two independent routes into the band converge on one record. The published route: three studies in leading peer reviewed venues describe the approach, the deployment and the clinician experience, with cohort sizes and settings disclosed, so an external reviewer can examine the methodology rather than take a claim on trust.

The company also publishes the specific claim that its approach produces fewer false alerts than prior systems, which is the metric alert fatigue actually turns on and the one competitors in monitoring almost never commit to in print.

The regulatory route: the sepsis module went through device clearance, with years of work validating performance across diverse hospital settings and patient populations, explicit assessment of missed cases and alert fatigue, and a post market monitoring and quality assurance programme.

That brings adverse event reporting, complaint handling and corrective action obligations that exist whether or not the vendor advertises them, and it gives a hospital a route that does not depend on the vendor granting one. What is still absent is the thing that separates B from A everywhere in this index: no warranty, indemnity, service level or remediation commitment attaching to the published performance was located. Ask whether the published false alert figures can go into the agreement, and for performance on the populations your hospital actually serves.

Integration is the delivery mechanism and it is unusually well substantiated. Models run inside the EHR in real time rather than in a parallel dashboard, and the founding team publicly reported partnering with both of the two largest electronic health record vendors specifically so the tool could be implemented at other hospitals. Reading live clinical notes plus structured data and delivering alerts into clinician workflow across every department, including the emergency department, is the deep end of this axis.

Converted from Not Rated. No hosting, tenancy, residency, retention or subprocessor terms were located, and the question the prior note raised has become sharper rather than softer.

That question was model isolation: the company describes adapting models per health system, which invites asking whether artefacts trained on one institution's data are separated from another's. That remains unanswered, and for a platform reading live records including clinical notes across every hospital department, it is the central deployment question.

What has changed is that per site adaptation now interacts with a regulatory constraint. The sepsis module received device clearance in April 2026. A cleared device is authorised with a defined intended use and defined technological characteristics, and a modification that could significantly affect safety or effectiveness generally requires a new submission. A buyer therefore needs to know exactly what per health system adaptation means for the cleared module: whether it is configuration within the cleared design, calibration of a threshold, or retraining that changes the model itself, and where the line sits between adaptation the clearance contemplates and modification that would take the deployed instance outside it.

That is a question about the vendor's regulatory operating model as much as its architecture, and it is one very few vendors in this index have to answer, because very few have crossed into device regulation while also promising per institution tuning.

Ask where the platform runs and under whose control, how one institution's model artefacts are isolated from another's, what per site adaptation changes, and how the company maintains the cleared configuration across sites that have each been adapted.

No published pricing or commercial structure. Enterprise health system contracting is the evident model given the deployment profile, but nothing about cost, per bed or per encounter basis, or implementation commitment is disclosed.

Acute inpatient care is the proven setting, with the published evidence spanning five hospitals across academic and community settings and every department including the emergency department. The company describes the platform as configurable to a wider array of condition specific use cases including clinical deterioration, pressure injuries, palliative care, transitions of care, and recovery at home. Buyers should note the crucial distinction: sepsis is the use case with published prospective evidence, and the others are stated platform capabilities without comparable published validation.