FDA Clears First AI Sepsis Detection System to Flag Patients Before Doctors Suspect It

Sepsis is one of the deadliest things that can happen inside a hospital, and one of the hardest to catch in time. It starts as an infection, then quietly tips into an overwhelming, body-wide immune response that can destroy organs within hours. By the time it looks like an emergency, it often already is one.

In May 2026, the U.S. Food and Drug Administration cleared an AI sepsis detection system built to close that gap. The Targeted Real-Time Early Warning System, known as TREWS, is now the first AI tool cleared to continuously monitor hospital patients for signs of sepsis before a clinician suspects it — not after. It’s an unusual case in health AI: a product with a genuine regulatory milestone, a published mortality benefit, and an active scientific debate about how much credit the algorithm deserves. All three are worth understanding.

The Problem: Sepsis Is the Leading Killer Hiding in Plain Sight

Sepsis is linked to roughly one in three deaths that occur in U.S. hospitals. At least 1.7 million U.S. adults and more than 18,000 children develop sepsis every year, and at least 350,000 adults and over 1,800 children die during that hospitalization, according to figures reported by CIDRAP, the infectious-disease news service run by the University of Minnesota (CIDRAP, 2026).

Time is the central problem. Research cited by the Johns Hopkins team behind TREWS has found that every hour treatment is delayed raises the risk of death by roughly 7 to 8 percent (Johns Hopkins Engineering Magazine). But sepsis rarely announces itself. Its early signs — a slightly elevated heart rate, a subtle drop in blood pressure, a lab value drifting out of range — can look like normal hospital noise until they suddenly don’t.

Hospitals have tried to solve this with rule-based screening tools and, more recently, machine learning. The most widely deployed AI example, the Epic Sepsis Model built into many hospitals’ electronic health record systems, was independently validated by University of Michigan researchers in JAMA Internal Medicine and found to have an accuracy (area under the curve) of just 0.63 — missing about two-thirds of sepsis cases while triggering alerts on 18 percent of all hospitalized patients (Healthcare IT News).

That track record is the backdrop against which TREWS’s FDA clearance matters.

Meet TREWS: From a Personal Loss to a Hopkins Lab to FDA Clearance

The algorithm behind TREWS was first published in 2015, when Johns Hopkins researchers Katharine Henry, David Hager, Peter Pronovost, and Suchi Saria described a “Targeted Real-time Early Warning Score” for septic shock in Science Translational Medicine (Science Translational Medicine, 2015). It was piloted in real hospital settings starting around 2017 at Howard County General Hospital, a Johns Hopkins-affiliated facility.

In 2018, Suchi Saria — a Stanford-trained machine learning researcher who holds the John C. Malone Endowed Chair at Johns Hopkins — founded Bayesian Health to commercialize the technology. Saria has said publicly that the work is personal: she lost a nephew to sepsis. Bayesian Health operates under a license agreement with Johns Hopkins University and, according to company and investor profiles, has raised more than $15 million from investors including Andreessen Horowitz, along with roughly $15 million in grant funding from the NIH, NSF, and DARPA (Bayesian Health).

Milestone timeline:

How TREWS Works

TREWS is not a physical device — it’s software that integrates with a hospital’s existing electronic health record system. It continuously reads data already flowing through a patient’s chart: vital signs, lab results, and clinical documentation. Using that stream, it computes an ongoing sepsis risk score for every monitored patient, from admission to discharge.

When a patient’s risk crosses a set threshold, TREWS surfaces an alert directly inside the clinician’s existing workflow — flagged on the patient list, and in some deployments escalated by paging or phone. A physician or nurse then reviews the alert and decides whether to act, typically by initiating the standard sepsis treatment bundle: blood cultures, lactate testing, fluids, and antibiotics (Nature Medicine, 2022).

The AI Technology Behind the Score

The exact architecture of Bayesian Health’s current commercial model is proprietary and undisclosed. What is known is that the original 2015 TREWScore used a regularized statistical model trained on bedside monitor and EHR data. The company today describes its product as an “Adaptive AI platform” tunable to different hospital populations and workflows, reasoning continuously across data types rather than firing on fixed, rule-based triggers.

It’s worth being precise about what “adaptive” means here. The Nature Medicine papers state plainly that “the underlying source code is proprietary intellectual property and is not available” — only the statistical analysis code behind the published outcomes study is public (Nature Medicine, 2022). FDA clearance, as a regulatory matter, only ever applies to a “locked” version of an algorithm — a fixed snapshot that can’t silently change after clearance. Any meaningful post-market updates to a genuinely adaptive system would, under current FDA practice, require additional review. That tension between “adaptive AI” language and “locked algorithm” regulatory reality recurs across clinical AI generally, not just at Bayesian.

The Evidence: What the Peer-Reviewed Data Shows

The strongest thing TREWS has going for it is unusual in health AI: a published mortality outcome, not just an accuracy score. Three companion studies appeared in Nature Medicine and Nature Digital Medicine on July 21, 2022, based on a prospective, multi-site study across five hospitals over roughly two years.

Those figures are genuine and peer-reviewed. They’re also the subject of real scientific disagreement about what they prove.

Independent Scrutiny: Why Some Researchers Urge Caution

Not everyone reads the 18.7 percent figure the same way. A peer-reviewed editorial in the American Journal of Respiratory and Critical Care Medicine raised a specific design concern: the comparison group — patients whose alerts weren’t confirmed within three hours — was “highly heterogeneous and may have included many patients without sepsis.” The same editorial noted that the study did not separately examine the more than 30,000 patients who received a TREWS alert but did not have sepsis, leaving open questions about false-alarm burden and potential downstream harm (AJRCCM).

A causal-inference critique posted to medRxiv, titled “TREWS or FALSE?”, argued that confounding factors — case complexity or distraction from false alarms — could plausibly explain part of the observed benefit rather than the early warning itself (medRxiv). When Prenosis introduced its competing Sepsis ImmunoScore in a paper published in NEJM AI, the authors noted that “recent reviews of validation studies of the Targeted Real-Time Early Warning System score have raised concerns regarding the control group and false positives” (NEJM AI).

The TREWS researchers have themselves acknowledged that a randomized controlled trial would strengthen the evidence, but reported difficulty operationalizing one in a live hospital setting. No such trial has been published to date.

Regulatory Status: What FDA Clearance Actually Means

TREWS received FDA Breakthrough Device Designation before 2023 — the exact date has not been publicly disclosed — allowing early deployment at several health systems ahead of full clearance. On May 12, 2026, the FDA granted 510(k) clearance to the platform, reported consistently by Johns Hopkins, CIDRAP, and trade outlets as the first clearance of an AI system specifically for continuous, pre-suspicion sepsis screening (Johns Hopkins Hub; Medical Device Network).

Some details remain unresolved publicly: the specific 510(k) submission number was not identified in available sources, and no CE mark or other non-U.S. regulatory clearance has been reported. Bayesian Health says the clearance supports its bid for reimbursement through Medicare and Medicaid’s New Technology Add-on Payment program, with a decision reportedly expected around August 2026 — not yet confirmed as of this writing.

Real-World Deployment: Who’s Using It Now

Beyond the original five-hospital study, TREWS is reported to be in use at Cleveland Clinic (across 20 hospitals in Ohio and Florida), MemorialCare, University of Rochester Medicine, Mayo Clinic, and MedStar, among others (CNBC, 2026). Bayesian has cited real-world results from the Cleveland Clinic rollout — a 46 percent increase in sepsis cases identified and seven times more cases caught early — but these figures come from the company and health system, not from independent peer-reviewed publication.

Benefits, Backed by Evidence

• Faster treatment: a median 1.85-hour reduction in time to first antibiotic order when clinicians engaged with alerts promptly.
• Reduced mortality, with caveats: an 18.7 percent adjusted relative reduction in in-hospital mortality among alerted, treated patients, per the 2022 study — an observational finding, not one from a randomized trial.
• High clinician engagement: 89 percent of alerts were evaluated by a physician or advanced practice provider.
• Earlier detection: company-cited data puts detection at 2 to 48 hours ahead of standard clinical recognition.

Limitations, Risks, and Open Questions

• No randomized controlled trial has yet validated the algorithm’s causal effect independent of confounding factors.
• False-positive burden is not fully characterized — the more than 30,000 non-sepsis alerts in the original study were not separately analyzed for potential harms.
• Model architecture and training data specifics are proprietary, limiting independent algorithmic audit.
• Generalizability beyond the original five participating hospitals is not fully established.
• Pricing and licensing terms are not publicly disclosed; the system is sold to health systems, not available to individual clinicians or patients.

The Competitive Landscape

What’s Next for AI Sepsis Detection

The most consequential near-term development is the pending Medicare/Medicaid reimbursement decision, which will materially affect how many hospitals can afford TREWS at scale. Longer term, the absence of a randomized controlled trial remains the clearest gap between what has been shown and what independent researchers say would be needed to settle the debate. Bayesian has also signaled ambitions to expand its platform beyond sepsis into other forms of early patient deterioration, including respiratory failure and cardiac events — a direction that will test whether the same evidentiary and regulatory questions apply more broadly.

Conclusion: A Genuine Milestone, With Real Caveats

TREWS’s FDA clearance is a legitimate first: the first AI system cleared specifically to watch for sepsis before a clinician suspects it, backed by a published, peer-reviewed mortality signal — something few clinical AI products can claim. But peer-reviewed is not the same as settled. Independent researchers have raised substantive concerns about how that mortality benefit was measured, and no randomized trial has yet resolved the question. For hospitals and patients, that leaves TREWS in a genuinely interesting place: a technology worth taking seriously, evaluated with the same rigor its own developers say it still needs.

FAQ

What is TREWS and what does it do?
TREWS is an AI system that integrates with a hospital’s electronic health record to continuously monitor patients and flag those at risk of sepsis, aiming to alert clinicians before they would otherwise suspect the condition.

Is TREWS FDA approved, and what does that clearance actually cover?
TREWS received FDA 510(k) clearance on May 12, 2026, for continuous AI-based sepsis monitoring, building on an earlier FDA Breakthrough Device Designation. The clearance covers the monitoring/alerting function — clinicians still review and act on every alert.

Does TREWS really reduce sepsis deaths — what does the evidence show?
A 2022 study in Nature Medicine, covering nearly 600,000 patients across five hospitals, found an 18.7 percent adjusted relative reduction in in-hospital mortality among sepsis patients whose alerts were confirmed within three hours. The finding is peer-reviewed but observational, and independent researchers have questioned aspects of the study’s control group.

What are the criticisms of the TREWS research?
Critics point to the comparison group used in the mortality study, which some researchers argue may not be a clean counterfactual, and to the lack of published analysis of the roughly 30,000 alerts issued for patients who did not have sepsis.

How is TREWS different from the Epic Sepsis Model or other sepsis AI tools?
TREWS is a standalone, FDA-cleared platform with published mortality outcomes, while the Epic Sepsis Model is built into the Epic EHR and was independently found to have weak predictive accuracy. Prenosis’s Sepsis ImmunoScore takes a different approach, incorporating blood biomarkers rather than relying solely on EHR vitals and labs.

Which hospitals currently use TREWS?
Reported users include Cleveland Clinic, MemorialCare, University of Rochester Medicine, Mayo Clinic, and MedStar, among other U.S. health systems.

How much does TREWS cost, and can patients or individual doctors access it directly?
Pricing has not been publicly disclosed. TREWS is sold to hospitals and health systems through institutional agreements and is not available directly to individual clinicians or patients.

This article discusses sepsis, a serious and potentially life-threatening medical condition. If you or someone you know is experiencing symptoms of sepsis — such as rapid heart rate, fever, confusion, or extreme pain or discomfort — seek emergency medical care immediately.

Editorial note: This article reflects publicly available information as of August 2026. Regulatory status, clinical trial results, and product availability for these tests may change. Readers should consult a licensed healthcare provider before making decisions.