Bayesian Health

The models are the product. Bayesian Health sells a platform of machine learning models that read the EHR continuously and flag patients at risk, with the cleared sepsis module as the lead and further modules for clinical deterioration, readmissions, palliative care, transitions of care, heart failure, pressure injuries and malnutrition.

The company describes a mixture of experts design (several specialized models and a routing layer that learns which to trust for each patient) and patient specific baselines, and the cleared sepsis device is an AI/ML algorithm reading 115 EHR parameters. Workflows that tee up order sets and track bundles sit on top of the flags, but nothing in the product works without the models.

How clinicians oversee each flag is described, but the risk threshold that produces one is not published. The cleared device raises a single flag, "Sepsis Risk High", in the EHR, for use with clinical assessment and laboratory data and not as the sole basis for deciding a patient has or will develop sepsis; each flag shows the clinical factors behind it. The company's published studies treated clinician response as part of the system: providers confirm or dismiss each alert, and the Nature Medicine outcome study measured outcomes by whether an alert was confirmed within three hours. On top of the flag, the company now describes agentic workflows that tee up order sets, pathway steps and documentation and track bundle completion, with the clinician acting.

Nothing describes what the workflow agents may do without a clinician's action beyond preparing it, or what happens when a flag is wrong.

The approach is described; the model sold is not identified. The platform page describes a mixture of experts design, with several specialized models and a routing layer that learns which to trust for each patient, patient specific baselines, and inputs drawn across the chart. The 510(k) summary says the cleared sepsis device uses 115 EHR parameters (history, comorbidities, chief complaint, labs, vitals, procedures, medications and consult orders), was tested against input noise and missing data, and has a post market plan for monitoring performance and communicating it to users. The company's published research describes the TREWS system and its methods for labeling sepsis in training data.

The deployed model carries no version name, and no published material characterizes the training data behind the cleared model. The company also does not say how and when site tuned models are updated, or how a site learns that its model changed.

Structurally short at the model layer for the lead product: the sepsis model is the company's own and is cleared as a specific device, so it cannot be swapped quietly for another provider's model. Everything around it is unnamed.

The device runs as a cloud web application, but no hosting provider, region or subprocessor list is published, and the agentic workflows the company now markets (preparing order sets and documentation, coordinating bundles) do not say whether any outside language model is involved.

The company's headline claim is lower mortality, and a prospective multi site outcome study answers it. Adams and colleagues (Nature Medicine, 2022) followed 590,736 patients monitored by the TREWS system at five Johns Hopkins hospitals, academic and community; among 6,877 sepsis patients flagged before antibiotics, those whose alert a provider confirmed within three hours had an adjusted 3.3 percentage point absolute (18.7 percent relative) reduction in in hospital mortality, and less organ failure and shorter stays, than those whose alert was not confirmed in time. Companion papers in Nature Medicine and npj Digital Medicine (2022) examined provider adoption and clinicians' experience. The company's own clearance study (K250680) adds a retrospective validation on 7,732 encounters at four hospitals.

What to weigh. The outcome study is observational: it compares patients by how quickly clinicians acted on the alert, not by whether the tool was used, so part of the difference may reflect the patients or teams rather than the tool. It was led by the company's founder at her academic institution and studied the system before it became the current cleared product. The company's summary page quotes an 18.2 percent relative reduction against the paper's 18.7 percent, and its home page cites a 5.5 percent absolute mortality reduction across customers and Cleveland Clinic figures (10 times fewer false alerts, 46 percent more identified sepsis cases) with no published method.

Only general assurances reach the product. The sepsis device's 510(k) summary says verification covered cybersecurity and data protection, and the device is a cloud based web application integrated with the EHR, reading 115 parameters including history, comorbidities, chief complaint, labs, vitals, procedures, medications and consult orders; the platform page says it ingests thousands of clinical inputs across the chart. The company's privacy policy covers its marketing website.

Nothing published states what patient data is retained, for how long, whether one health system's data develops models used by others, or who inside the company can reach identified records. The per site adaptation the company describes makes the training question concrete, since a model tuned to a hospital is built from that hospital's records.

No statement of business associate or HIPAA status, and no privacy document that reaches the product. The company's privacy policy covers its marketing website (cookies, newsletter sign ups, surveys). A platform reading live EHR data at health systems necessarily operates under business associate agreements, but nothing about them is published. Device clearance does not answer this: FDA examines safety, effectiveness and cybersecurity design, not the terms on which a vendor may use patient information.

An outside check of the software, and no attestation of the company's operations. The sepsis device's 510(k) (K250680) included a cybersecurity risk analysis and testing under FDA's premarket and post market cybersecurity guidance, reviewed by FDA as part of clearance. That examines the device's security design, not how the company runs its own environment, and it is not published. No SOC 2 report, ISO 27001 certificate, HITRUST certification, trust center or request path is published, and nothing covers the wider platform beyond the cleared module.

Cleared and marketed inside its indication. K250680 (Bayesian Health Sepsis Flagging Device, decided 30 April 2026; 21 CFR 880.6316, software device to aid in the prediction or diagnosis of sepsis, Class II, product code SAK; predicate Prenosis Sepsis ImmunoScore, DEN230036) clears a continuously monitoring AI/ML device that flags adult patients (18 and over) at high risk of sepsis within 24 hours, from emergency department presentation or admission through the stay in acute care, as an aid used with clinical assessment and not as the sole basis for a decision; prescription use. The company's pages attach the cleared label to its sepsis monitor, which matches, and the device previously held breakthrough designation.

The clearance covers the sepsis module only; the deterioration, readmission, palliative care, transitions, heart failure, pressure injury and malnutrition modules are marketed without a cleared label and without a stated regulatory basis. The clearance summary records that the clinical study missed its pre-specified sensitivity criterion and that clearance rested on the totality of evidence. No change control plan is described, which matters because the company tunes models per site.

A governance process is described and a subgroup finding is stated, without published figures. The company's platform page describes a built in governance layer that tracks real world performance and detects drift and bias across populations and sites.

The 510(k) summary for the sepsis device (K250680) reports that its clinical validation found no statistically significant differences in performance across race, gender, age and ethnicity, and describes a post market performance management plan covering data collection, monitoring, corrective action and communicating performance to users. No subgroup results are published as numbers, so a buyer cannot see where the model does less well. The company's argument that one size fits all models fail across different populations and care patterns is the motive for its per site approach.

The company has put an error rate on the record in its own regulatory filing. The 510(k) summary for the sepsis flagging device (K250680, April 2026) reports, on 7,732 encounters at four hospitals, retrospective adjudication by physicians: encounter level positive percent agreement 79.4 percent (185 of 233 septic encounters flagged within 24 hours of onset), negative percent agreement 89.5 percent (6,715 of 7,499), and a flag level positive predictive value of 11.7 percent at a sepsis prevalence of about 3 percent. The summary states plainly that the device did not meet its pre-specified acceptance criterion for positive percent agreement (a lower confidence bound above 80 percent) and was cleared on the totality of evidence. So about one septic patient in five was not flagged within the 24 hours around onset, and about nine flags in ten were not a true positive in that window.

Nothing is published about what the company answers for when the output is wrong: no warranty, indemnity or remediation commitment.

Integration is described in depth and no record system is named. The 510(k) summary calls the device's web application EHR agnostic and integrated into the EHR so clinicians see flags in their existing workflow, and the platform page says flags trigger workflows inside the EHR that tee up order sets, pathway steps and documentation.

But the company names no EHR vendor, interface standard, marketplace listing or integration document; Epic appears on its site only as the source of a comparison with a customer's previous sepsis algorithm. Named deployments at Johns Hopkins and Cleveland Clinic show the integration runs in production.

A single hosted option with the location not stated. The 510(k) summary describes the sepsis device as a cloud based web application integrated into the EHR; no hosting provider, region, tenancy model, residency commitment or retention term is published. Named deployments include Johns Hopkins and Cleveland Clinic, where the sepsis module is live enterprise wide.

The company adapts models per health system, and for the cleared module that raises a question the company does not answer publicly: whether per site adaptation is configuration within the cleared design or retraining that changes the model, and how site specific models are isolated from one another.

No price is published and the posture is discoverable. The product is sold to health systems as an enterprise platform, by module, with implementation that tunes models to each site. The company announced in September 2026 that its cleared sepsis device was approved for a Medicare New Technology Add-on Payment from 1 October 2026: hospitals can bill up to $61.84 per eligible adult Medicare fee for service inpatient case monitored, across an estimated 739 MS-DRGs, for up to three years, and it publishes a reimbursement guide. That is what Medicare pays the hospital, not the company's price.

Coverage is named with validation behind part of it. The cleared sepsis module covers adults 18 and over from emergency department presentation or admission through the stay in acute care, including observation units, general wards and intensive care, and its evidence spans five hospitals in the outcome study and four in the clearance study, academic, urban and suburban.

The platform is also marketed for clinical deterioration, readmissions, palliative care, transitions of care, heart failure therapy, pressure injuries and malnutrition, and the company has presented a palliative care program built with Mayo Clinic, but none of those modules has published validation comparable to sepsis.