FDA Clears First AI That Watches Every Patient for Sepsis, Cuts Deaths 18%

The First Device That Watches Every Patient

The FDA has never cleared a device that watches every hospitalized adult for sepsis without a clinician asking it to. Until now, every approved tool required a human to suspect sepsis first: order a blood culture, trigger a risk score, initiate the workflow. Bayesian Health's Sepsis Flagging Device (K250680) breaks that pattern. It runs continuously from emergency department presentation through discharge, reading the full electronic record to establish each patient's baseline and flagging deterioration in people who don't yet look sick.

The clearance came through the traditional 510(k) pathway on April 30, 2026, after a 420-day review. The predicate device is Prenosis's Sepsis ImmunoScore (DEN230036), cleared April 2024. But the Bayesian device differs in three ways the FDA summary highlights: continuous monitoring, independence from blood culture orders, and applicability to all adult inpatients regardless of clinical suspicion. The agency classified it under product code SAK, regulated by the Gastroenterology & Urology panel under 21 CFR 880.6316. No predetermined change control plan was authorized.

A 2022 Nature Medicine study anchored the submission: 764,707 patient encounters (17,538 with sepsis) across five hospitals in academic and community settings, with more than 2,000 providers using the software. The research showed 82% sensitivity, a 5.7-hour median lead time over standard detection, and 89% provider adoption. One in three flagged cases was physician-confirmed. Encounter-level positive percent agreement reached 79.4%; negative percent agreement hit 89.5%. Flag-level positive predictive value at the observed 3% sepsis prevalence was 11.7%. The device missed its pre-specified encounter-level PPA target, but the FDA determined the totality of performance (including flag-level metrics) established substantial equivalence.

When clinicians acted on the alerts in time, sepsis patients were 18% less likely to die in the hospital, the PRNewswire release found. That relative mortality reduction, 18.2%, is the figure Bayesian cites most often. Sepsis drives more than $50 billion in annual U.S. hospital costs, PRNewswire reported. Each hour of delayed treatment drops survival by up to 7.6%. The platform outputs a single "Sepsis Risk High" flag inside the EHR — not a diagnosis, not a standalone decision tool. The label warns it should not be used as the sole basis for determining sepsis presence or 24-hour risk.

"Deploying AI for sepsis detection across a health system like Cleveland Clinic requires rigorous evidence and seamless clinical integration," said James Morrison, MD, chair of the Cleveland Clinic Enterprise Sepsis Steering Committee. Cleveland Clinic is rolling the tool out across its Ohio and Florida locations. At University of Rochester Medicine, Chief Digital and Information Officer Gregg Nicandri, MD, said clinicians found Bayesian provides early detection with a high rate of accuracy, a tool they can trust and use to diagnose and treat patients.

A post-market performance management plan will monitor the device over its market lifetime. The clearance also positions Bayesian for potential New Technology Add-on Payment approval from CMS starting October 2026, with a final decision expected early August. The sepsis flagging device is one module of a broader clinical intelligence platform designed to address the leading drivers of deterioration, mortality, length of stay, and ICU utilization through a single, continuously learning system. The question now is whether health systems can operationalize that intelligence at the bedside without drowning clinicians in another alert stream.

Inside the EHR, Not Beside It

Bayesian Health's platform does not sit beside the EHR. It runs inside it. The company built its sepsis monitor on SMART on FHIR, the open framework that lets third-party apps launch natively from within Epic and Cerner charts using OAuth2 and OpenID Connect. When a clinician opens a patient record, the Bayesian module appears as a contextual panel: no separate login, no context switch, no new tab. The app receives the patient identifier, encounter ID, and user role automatically at launch, then pulls the live chart data it needs through FHIR R4 resources: observations, medications, procedures, encounters, and clinical notes.

Epic and Cerner together power the vast majority of large U.S. health systems, and both have committed to FHIR R4. Epic's App Orchard now lists 790 SMART apps, 344 added since 2024, signaling rapid ecosystem adoption. Cerner, now Oracle Health, deprecated its legacy DSTU2 interfaces by December 2025, making R4 the only supported standard. Bayesian targets both vendors natively. Its SMART app authenticates via Epic's strict SMART App Launch flow or Cerner's Well-Known SMART Configuration for endpoint discovery, handling each system's token refresh mechanics and rate limits in production.

But a SMART app alone cannot sustain continuous, hospital-wide surveillance. The specification constrains data access to the active user session. To monitor every patient around the clock (across the five-hospital validation cohort), Bayesian pairs the in-workflow app with a backend integration layer. That backend runs on the company's infrastructure, pulling longitudinal data through bulk FHIR exports, HL7 v2 feeds, and scheduled FHIR queries. It normalizes the raw JSON resources (inconsistent coding systems, optional fields, vendor-specific extensions) into structured, ML-ready feature stores. The backend also handles the computational load of continuous risk scoring across thousands of simultaneous patients, something a session-bound app could never sustain.

The architecture mirrors what integration veterans now recommend: build the backend foundation first for data depth and AI readiness, then layer the SMART app for clinician-facing workflows. Bayesian's SMART component surfaces the risk score, the key contributing factors, and the recommended action set directly in the chart. The backend feeds it continuously. Clinicians see a single, prioritized alert only when the model crosses a calibrated threshold — one in three flagged cases were physician-confirmed in the study — rather than the noisy, one-size-fits-all warnings that drive alert fatigue.

This design produced 89% physician adoption across 2,000+ providers. The platform "feels like a native part of Epic or Cerner," as one integration analysis put it, because it respects the EHR's operational realities: short-lived tokens, paging limits on deep history, partial note-text access, and constrained write-back (observations, documents, tasks, care-plan entries). Bayesian does not assume the FHIR spec translates uniformly across sites; it accounts for site-specific scopes, governance policies, and extension handling. The result is a system that delivers real-time risk stratification and triage to front-line clinicians, in-workflow, without demanding new habits or new screens.

That architecture demands a rare engineering profile: fluency in SMART on FHIR, OAuth2 flows, vendor-specific sandbox quirks, and the write-back patterns that let an AI insight surface inside a nurse's existing Epic in-basket. Bayesian's career page lists 11 open roles, all remote, five in engineering, including a Clinical Integration Nurse Consultant who translates workflow requirements into technical specs. The company's hiring mirrors the dual-competency model health systems themselves are building: a technical integration lead paired with a clinical translation lead. Engineers who combine interoperability standards, regulatory awareness, and clinical workflow empathy are scarce; the remote model expands the pool beyond traditional hubs but raises the bar for asynchronous delivery during critical go-live periods.

Outcomes That Move the CFO and CMO

The Nature Medicine findings translate into the language of the C-suite. Across the five-hospital prospective deployment running 2.5 years with over 4,000 caregivers using the system in real time across every department including the emergency department, Bayesian's TREWS platform delivered the mortality reduction above when providers acted on alerts within three hours. Earlier intervention drives the signal: providers who confirmed an alert within three hours ordered first antibiotics 1.85 hours faster than those who dismissed or delayed.

At MemorialCare's Long Beach Medical Center, a four-month assessment after go-live showed time to antibiotic administration cut by more than half when clinicians engaged the Bayesian flag. "When providers engage with the Bayesian flag, we're seeing a 3.6 percent absolute mortality reduction, and time to antibiotics is cut in half when they engage within the first hour," said Dr. Leo, MemorialCare's physician lead. ED adoption there hit 90 percent.

Adoption is the lever. The studies recorded 89 percent of all alerts evaluated by a physician or advanced practice provider, and 38 percent of evaluated alerts confirmed (roughly one in three). That confirmation rate matters: patients whose alerts were confirmed within three hours captured the mortality and organ-failure benefits; those whose alerts were dismissed or delayed did not. Contrast that with MemorialCare's prior experience on Epic's Sepsis Predictive Analytics v.1. That tool "had limited sensitivity and specificity, which led to significant over-alerting and alert fatigue and did not result in a measurable reduction in sepsis mortality," per the health system's own 2026 retrospective. Bayesian's design difference is structural: a passive flag embedded in the Epic workflow that asks the clinician whether they agree, then guides assessment per Surviving Sepsis Campaign guidelines if they do. The model learns from each response, continuously recalibrating. Clinicians at MemorialCare now work with "significantly fewer electronic alerts," which Dr. Leo credits for restoring trust and sustaining engagement.

Health systems measure ROI in three columns: lives saved, length-of-stay avoided, and alert-fatigue cost recovered. The studies show reductions in organ failure and length of stay alongside mortality. MemorialCare is expanding to Orange Coast Medical Center and Saddleback Medical Center using the same physician-champion model that worked at Long Beach. Cleveland Clinic, Johns Hopkins Health System, and University of Rochester Medicine are also live. The common thread: each system treated the deployment as a clinical program with local validation, not a software install. The FDA clearance (K250680) established the floor; the workflow integration and measured adoption built the ceiling. That adoption reality sets the bar for every other clinical AI vendor now pitching health systems.

Where Bayesian Stands in Clinical AI

Bayesian Health's FDA clearance arrives in a market where every major EHR vendor and cloud platform now markets an "AI layer." The distinction matters: most of those layers optimize documentation or revenue cycle. Bayesian's clearance is for a clinical decision support device that continuously monitors every hospitalized patient for sepsis — no prior suspicion required — and has peer-reviewed evidence that clinicians actually use it.

Epic controls roughly 42% of the U.S. acute-care market and has embedded generative AI across its suite. Art for Clinicians drafts end-of-shift notes; nurses write notes 85% faster, Epic reports. Emmie for Patients handles patient-facing chat. Penny for Revenue Cycle targets denials. These are real productivity gains, but they sit on top of the EHR's episodic data model, triggered by a clinician opening a chart or a coder reviewing a claim. They do not continuously reason over vital signs, labs, and notes across every patient in the background. Epic's AI roadmap centers on generative AI and cognitive computing with partnerships including Abridge for documentation and Microsoft for infrastructure.

Cerner, now Oracle Health with 22% market share after Oracle's $28.3 billion acquisition, emphasizes accessibility and the CommonWell Health Alliance for data exchange. Its systems are "easier to integrate from a complexity perspective because they offer fewer features than Epic systems." But like Epic, its native AI efforts center on workflow automation and interoperability, not continuous clinical surveillance validated for mortality reduction.

Microsoft's Azure Health Bot provides a compliant conversational AI framework — "built-in medical intelligence with natural language capabilities, extensibility tools and compliance constructs" — that health systems use to build symptom checkers, triage chatbots, and patient-navigation tools. It is a platform, not a clinical application. Health systems still need to design the clinical logic, validate it, and integrate it into workflows. The same holds for Microsoft's broader "unify healthcare data in solutions powered by AI to create a 360-degree view of patient health" vision: the integration and validation burden remains with the provider.

Palantir Foundry takes a different angle. Its AI/ML capabilities "empower health care systems by securely integrating and modeling diagnosis, clinical, and purchasing data, resulting in a clinically-integrated supply chain solution that can respond to changes in real time." The platform excels at operational analytics (supply chain, capacity planning, cohort identification) and Palantir's hiring reflects that focus (10 roles added in the past week, median board salary $160k across 233 salaried roles), Zero G Talent's board data shows. But Foundry is a general-purpose data operating system. Building a continuous sepsis monitor on top of it means the health system owns the regulatory submission, the clinical validation, and the workflow integration.

Bayesian's differentiation rests on three grounded factors. First, the regulatory bar: it is the first and only FDA 510(k)-cleared device for continuous AI sepsis monitoring across all hospitalized patients. Other FDA-cleared sepsis solutions (biomarker blood tests or risk scores calculated from those tests) require a clinician to suspect sepsis and order the test. Bayesian flags deterioration before clinical suspicion, running continuously on the 2–4% of acute-care patients who will develop sepsis, not the 20–40% already suspected.

Second, the evidence base: the five-hospital study with 2,000+ providers showed adoption at 89% — a rate "rarely seen in clinical AI, where most tools are ignored or turned off." Real-world deployments across Bayesian customers show 81% adoption, 3.3% absolute mortality reduction, and half a day length-of-stay reduction. A health system CMO quoted on Bayesian's site put it bluntly: "We tried multiple tools, all with less than 1% adoption. At that rate, you can't improve anything, no matter how good the model is. Bayesian fires 3 hours earlier and 20 times less often on patients who are actually getting sick."

Third, the architecture: Bayesian built on SMART on FHIR to embed inside Epic and Cerner workflows, not beside them. The platform "continuously reads the full patient record to establish each patient's baseline and detect meaningful change over time," applying "complex clinical reasoning to identify the patients who truly need attention and surface clear next steps." It adds a continuous intelligence layer without requiring clinicians to leave the EHR, open a separate dashboard, or interpret raw model outputs.

The competitive implication is clear. Health systems evaluating AI for clinical deterioration face a build-versus-buy choice framed by regulatory risk and adoption reality. Epic and Oracle Health will keep expanding native AI for documentation and operations. Microsoft and Palantir sell platforms that health systems must customize, validate, and maintain. Bayesian delivers a cleared, validated, workflow-embedded application for the highest-mortality condition in U.S. hospitals — with the adoption data to prove clinicians trust it.

The Engineer Who Speaks Clinician

Health systems building internal AI operations teams now hunt the same profile, engineers who speak FHIR, understand 510(k) pathways, and translate clinical workflow into technical specs. The talent market hasn't caught up. But the standard has been set: a cleared device, a validated workflow, and a single alert that clinicians actually trust.

In the end, the breakthrough isn't the model — it's the flag. A single "Sepsis Risk High" banner appearing in a clinician's workflow, backed by an 18% mortality reduction when acted on. The FDA has cleared the device; the evidence has validated the workflow; the health systems are deploying. The question now is whether the rest of clinical AI can meet this standard: cleared, continuous, and actually used.

Working in frontier tech? Zero G Talent tracks the openings: see every open Palantir Technologies role, browse frontier tech jobs, the companies hiring, and the people building the field.