Sepsis AI linked to 18.7% lower mortality wins FDA clearance and Medicare add-on payment

Bayesian Health's sepsis flagging device, the commercial form of the Targeted Real-time Early Warning System (TREWS) developed at Johns Hopkins University, received FDA 510(k) clearance on 30 April 2026 and was approved for a Medicare new technology add-on payment in the FY 2027 inpatient payment final rule published on 4 August 2026 ¹⁴⁶. The published outcome evidence for the underlying system dates from 21 July 2022, when Nature Medicine carried two papers on a deployment of TREWS across five hospitals, with Roy Adams, Katharine Henry and Suchi Saria among the authors ¹².

What the studies found

The outcomes paper by Adams and colleagues was a prospective, multi-site cohort study. TREWS monitored 590,736 patients across the five hospitals, and the analysis focused on 6,877 patients with sepsis whose alert fired before antibiotics were started ¹. After adjustment for presentation and severity, patients whose alert was confirmed by a provider within 3 hours had an adjusted absolute reduction in in-hospital mortality of 3.3% (CI 1.7 to 5.1%) and an adjusted relative reduction of 18.7% (CI 9.4 to 27.0%) compared with those whose alert was not confirmed within 3 hours ¹. Among patients also flagged as high risk, the absolute mortality reduction was 4.5% (CI 0.8 to 8.3%) ¹.

The companion paper by Henry and colleagues covered a 2-year period. Among 9,805 retrospectively identified sepsis cases, TREWS identified 82%. Providers evaluated 89% of alerts and confirmed 38% of evaluated alerts. Patients whose alert was confirmed within 3 hours had a 1.85-hour reduction (95% CI 1.66 to 2.00) in median time to first antibiotic order ². A third paper in npj Digital Medicine, based on interviews with 20 physicians and nurses at a community hospital, found that clinicians "perceived themselves as partnering with the technology" ³.

The 18.7% figure describes the gap between sepsis patients whose alert was acted on quickly and those whose alert was not, within a cohort design. It is not a comparison of hospitals with and without the tool. Bayesian Health's own release gives the figure as 18.2% ⁷. The authors disclosed that Johns Hopkins University and two of them are entitled to revenue under a licence agreement with Bayesian Health, that the university owns equity in the company, and that Saria is its founder ¹³.

What it means for the market

The FDA cleared the device as a Class II software device under 21 CFR 880.6316, with Prenosis's Sepsis ImmunoScore as the predicate. The indication is to aid early detection or risk prediction of sepsis developing within 24 hours in adults, from emergency department presentation through the acute stay ⁴. Bayesian Health said no continuous AI sepsis monitor had been cleared before ⁵. CMS then finalised a maximum add-on payment of $61.84 per case, against a cost per case of $95.14, identified by ICD-10-PCS code XEZZXJC, with the newness period starting on 30 April 2026 ⁶. The company says hospitals can bill from 1 October 2026 across an estimated 739 MS-DRGs, for up to three years ⁷.

The FDA summary gives buyers a more sober picture than the 2022 headline. In the clinical validation of 7,732 encounters from 7,298 patients at 4 hospitals, encounter-level positive percent agreement was 79.4% (95% CI 74.2 to 84.6) and negative percent agreement 89.5% (95% CI 88.8 to 90.2). Flag-level positive predictive value at the observed sepsis prevalence of about 3% was 11.7% (95% CI 10.0 to 13.2), and the device did not meet its pre-specified acceptance criterion for encounter-level PPA ⁴. One commenter questioned the device's performance on the basis of that summary; CMS replied that performance concerns are not within the scope of its review under the alternative pathway for Breakthrough Devices ⁶.

What remains unknown is whether the 2022 mortality association holds elsewhere. Bayesian Health says the results "have since been matched or exceeded across a diverse set of health systems", but its release cites no peer-reviewed outcome study beyond the five hospitals in the 2022 papers ⁷.

The FDA notes the device has not been validated in paediatric, labour or maternity patients, patients who had surgery in the past 90 minutes, or trauma patients who have not been stabilised, and it is covered by a post-market performance management plan ⁴. A commenter asked CMS to monitor false positive and false negative rates across patient populations so that algorithmic tools do not widen disparities in sepsis recognition ⁶.

The add-on payment applies to FY 2027 discharges from 1 October 2026, at a maximum of $61.84 per case ⁶⁷.