Prenosis

The product sold to hospitals is the model. The De Novo decision summary for the Sepsis ImmunoScore (DEN230036) describes a locked probability random forest of 1,000 extremely randomized trees, calibrated by logistic regression, that turns up to 22 record and laboratory inputs into a risk score, one of four risk categories and a per-input explanation, delivered from a cloud service into the record or a web view.

There is no instrument, monitor or separate workflow platform behind it: the laboratory tests it reads (procalcitonin and C-reactive protein among them) are ordinary hospital tests, and the display is the model's output. The company's Immunix biobank and research platform, more than 138,000 specimens from over 35,000 patients across 11 hospitals by its own count, is how the model was built and is not what a hospital buys.

The oversight structure is described, with the thresholds published, and the action expected at each level is not. The score must be ordered by a clinician like a laboratory test, and the product page states it is not an alert system. The indication (DEN230036) says it is to be used with other laboratory findings and clinical assessment and should not be the sole basis for deciding that a patient has or will develop sepsis.

The decision summary publishes the category cut points (low below 12.2, medium to 30.6, high to 87.2, very high above) and how they were set, including a low threshold tuned for sensitivity at a 50 percent false positive rate, and each result shows how each of the 22 inputs raised or lowered the risk. What is not published is what a clinician is expected to do in each category, how the score relates to sepsis bundle timing, or how a result later shown to be wrong is fed back.

The model and its training data are described; version and update practice are not. The FDA decision summary (DEN230036) names a probability random forest of 1,000 extremely randomized trees (minimum node size 13, eight candidate variables per split) with Platt calibration, bag imputation for missing optional inputs, SHAP values shown with each result, and training on 2,366 patients from three sites of the company's NOSIS dataset labeled by encoded Sepsis-3 criteria and physician chart review.

It states the algorithm is locked. The company's JAMIA Open (2026) paper describes the SHAP display. No version identifier, release history or statement of how a model change would be made and communicated is published, so a buyer cannot tell which model version is running.

The chain is structurally short at the model layer. The model is the company's own, trained on its biobank and partner hospital data and locked under the De Novo authorization (DEN230036), so it cannot be swapped without a regulatory change. The other parties are partly named: Roche distributes the score through its navify Algorithm Suite, Vanta supports the company's SOC 2 program and Advantage Partners performed the audit. The cloud host, any subprocessors and the terms that govern hospital data once it arrives are not published.

Independent site validation reviewed by FDA and a peer reviewed multicenter study, with the method readable, and the results are good discrimination with misses at the low end. The De Novo validation (DEN230036) used prospectively collected specimens from three sites not used in training, physician adjudication blinded to the score and a check for adjudication bias; the AUROC was 0.81 (forced majority) and 0.84 (unanimous), sepsis rates rose across the four categories as prespecified, and 7 of 164 septic patients fell into the low category.

Watson et al. (Diagnostics, 2026; 6,027 patients, seven US hospitals, a retrospective analysis of prospectively collected samples) report AUC 0.82 for sepsis against 0.72 for SOFA and 0.70 for procalcitonin, 0.80 for in-hospital mortality and 0.74 for ICU admission. Company staff lead both papers, with hospital investigators as coauthors. A BARDA funded randomized trial of 800 patients with severe respiratory infection is announced, with no results yet.

General assurances, and a data flow described by the regulator rather than stewardship terms from the company. The FDA decision summary (DEN230036) says the device is a cloud-based system that takes up to 22 values from the hospital record as they are measured and stores the score and its components; the company's April 2026 SOC 2 release says its controls protect the clinical and patient data central to Immunix and the ImmunoScore.

The Diagnostics (2026) paper states that the specimen study behind the model was approved by each hospital's ethics board under a waiver of consent. Nothing published states how long hospital record data and scores are retained, whether customer data is used to build or retrain models, or how it is de-identified, and the privacy policy (June 2023) addresses website and subscription data only. Nothing is published on safety engineering beyond the regulatory controls, such as how a scoring fault in the field is detected and handled.

No business associate status is stated and no privacy document reaches the service that handles patient data. The privacy policy (last updated June 2023) names the company's software platforms in its scope but describes only website visitors, subscribers, cookies and marketing contact data, and says nothing about protected health information, hospital data or patients.

No business associate agreement, statement of status or HIPAA position appears on the product, solutions or evidence pages, in the privacy policy or in the company's list of news releases. The company's SOC 2 release (April 2026) describes security controls for clinical and patient data, which is security rather than a privacy position. Sales run directly and through Roche, and a hospital's data may pass to either.

A SOC 2 Type II report is named with its auditor and date, without the report. The company's April 2026 release says Advantage Partners completed the audit, supported by Vanta, with the report issued on 27 March 2026 and annual recertification; no trust center or copy of the report is linked.

The FDA decision summary (DEN230036) records that the cybersecurity information the agency's premarket guidance asks for was submitted, including a threat model, software bill of materials, mitigation of adversarial examples, cyber testing and a plan for postmarket vulnerabilities, an outside review of the device software. No ISO/IEC 27001 or HITRUST certification or penetration testing cadence is published.

A De Novo authorization identifiable in the FDA database, covering the model, with its indication stated. DEN230036 (granted 2 April 2024, product code SAK) authorizes the Sepsis ImmunoScore as AI/ML software that uses up to 22 inputs from the record to give a risk score and one of four categories, to aid in assessing the risk of presence of or progression to sepsis within 24 hours for emergency department or hospital patients with suspected sepsis and a blood culture ordered, not as the sole basis for that decision, prescription use only.

It created a new classification with special controls. The company's pages describe it as authorized for the diagnosis of sepsis and the prediction of adverse outcomes; the indication is sepsis risk within 24 hours, and the categories' associations with mortality, ICU transfer, ventilation, vasopressors and length of stay are reported in the summary as secondary endpoints.

Per group results are published with a stated method and the weaker groups are named. Watson and colleagues (Diagnostics, 2026; Prenosis staff with investigators at seven US hospitals; 6,027 adults with suspected infection, 25.8 percent Black and 7.7 percent Hispanic), listed on the company's evidence page, tested discrimination within subgroups of sex, age and race and ethnicity with bootstrap confidence intervals and between-group tests.

For sepsis the AUC ranged from 0.80 to 0.83 across subgroups; for secondary outcomes it was higher in women than men for in-hospital mortality (0.83 against 0.77) and higher in younger than older patients for mortality (0.84 against 0.76) and ICU admission (0.76 against 0.72), and did not differ by race or ethnicity.

The FDA decision summary (DEN230036) records that subgroup analyses by age, sex, race, immunocompromised status, imputation and site were submitted, without printing them, and the classification's special controls require labeling to name subpopulations with disparate performance. The company's Responsible AI framework is offered as an eBook behind a form.

A published error profile with case counts, adopted by the company's citation of its De Novo grant, and nothing standing behind the output. The FDA decision summary (DEN230036, April 2024) reports the locked model on a validation cohort from three sites independent of training (Beth Israel Deaconess, Jesse Brown VA, Beaumont Royal Oak; forced majority adjudication, 698 scored patients): sepsis in 7 of 232 low category patients (3.0 percent), 20 of 157 medium (12.7 percent), 101 of 276 high (36.6 percent) and 23 of 33 very high (69.7 percent), with likelihood ratios from 0.11 to 8.33 and an AUROC of 0.81.

The summary names the concern that septic patients assigned the low category could be treated late, and reports their outcomes as less severe. These are test set results, not performance in use. No warranty, indemnity or remediation commitment attached to a result is published, and no route for a patient or clinician to challenge one.

Record integration is described, with a named deployment behind it and no record system named. The product pages say the score integrates directly into hospital records, reading the 22 inputs from routine data and displaying the result with its explanation in the clinician's workflow, and the FDA decision summary notes that the validation sites used different record systems.

A July 2026 news item reports LifeBridge Health's Sinai Hospital of Baltimore putting the score into clinical use, and Roche's navify Algorithm Suite is named as a distribution route that connects to hospital laboratory and record infrastructure. No record vendor, interface standard or integration documentation is published.

A single hosted option with location implied rather than committed. The FDA decision summary (DEN230036) describes a cloud-based system that receives record data as it is measured, stores the score and its components, and shows the result in the record or through a web interface; distribution through Roche's navify Algorithm Suite is a second route to the same service. No hosting provider, region, tenancy or retention for stored scores is published, and the privacy policy notes transfers to other countries for website data only.

No price is published, and how the product is sold and paid for is discoverable. The company sells directly with integration into hospital records and through an exclusive distribution collaboration with Roche on the navify Algorithm Suite (announced April 2024), and publishes a hospital billing and coding guide.

Its October 2025 release explains that CMS created ICD-10-PCS code XXEZXCB (New Technology Group 11) to document use of the score, a prerequisite for a New Technology Add-on Payment application; that is documentation and a payer route, not the company's price.

The setting is narrow and it is the one validated. The indication (DEN230036) is adults in the emergency department or hospital with suspected sepsis and a blood culture ordered, and both the FDA validation (three independent sites) and the Diagnostics (2026) study (seven US hospitals, adults with suspected infection) were run in that population. It is sold as a diagnostic that combines laboratory biomarkers with record data. Pediatric, outpatient and other conditions are outside the indication, and the BARDA funded trial in severe respiratory infection is where the company is testing a wider use.