Insights from clinical leaders at Cleveland Clinic, MemorialCare Health System, and University of Rochester Medical Center

Sepsis kills more than 270,000 Americans every year. Most health systems know this. Most have invested in detection tools. And most will privately admit those tools haven't delivered.

Alert fatigue. Provider disengagement. Mortality curves that flatline after years of effort. It's a story that sounds familiar to any clinician who has ever silenced a sepsis alert and moved on.

At a recent Bayesian Health webinar, three clinical leaders who have moved past that story shared what it actually takes to operationalize AI-driven sepsis detection — not pilot it, not evaluate it, but run it live across complex health systems and see real outcomes. Their institutions are MemorialCare Health System, University of Rochester Medical Center (URMC), and Cleveland Clinic.

What follows are the key lessons, stats, and hard-won insights from that conversation.

Why Legacy Alerting Was No Longer Good Enough

Each panelist arrived at Bayesian Health through a version of the same breaking point: a detection tool that generated noise faster than it generated trust.

At MemorialCare, Dr. Jim Leo — CMO Emeritus and the architect of the system's sepsis program since 2008 — had already driven mortality down by 75% through years of disciplined improvement work. Then progress stalled. "We plateaued despite the use of Epic's version one predictive analytics," Dr. Leo explained. "Those analytics were producing tremendous over-alerting — a lot of alert fatigue, both poor sensitivity and poor specificity."

At URMC, the situation was starker. Dr. Conrad Gleber, Associate CMIO, described a provider adoption rate that had collapsed to nearly zero: "We had a response rate — an adoption rate — for providers that was less than 1%. So despite the alert going off, no one was engaging with it."

At Cleveland Clinic, Dr. James Morrison — an emergency physician, intensivist, and Chair of the Enterprise Sepsis Committee — framed the problem in terms of what detection tools need to do that rule-based systems simply can't: find patients earlier, with less noise, in a way that moves clinicians to act.

The common thread: the problem wasn't that clinicians didn't care about sepsis. It was that the tools had burned through their credibility.

The Validation Process: What Rigorous Independent Testing Looked Like

Each institution independently validated Bayesian Health's model before going live, and each approached it with genuine scrutiny.

URMC: Shadow Mode Testing Against Real Outcomes

URMC ran Bayesian Health in the background against their live patient population, measuring it against Epic, which had been implemented in 2017. The question wasn't just whether the model performed better in the abstract — it was whether it would perform well enough in their specific population to earn clinical buy-in.

The results that mattered most:

• 20x reduction in alert volume compared to the previous system

• Approximately 3 hours of additional lead time before the prior alert would have fired

"It does matter to a bedside clinician that it's going to alert you three hours earlier than you would normally have been alerted, and 20 fewer times per patient," Dr. Gleber noted.

MemorialCare: Head-to-Head Against Epic V2

MemorialCare ran a direct comparison between Bayesian Health and Epic's sepsis model version 2. The outcome: Bayesian Health delivered more than double the sensitivity with meaningfully better positive predictive value. More sepsis patients caught. Fewer false alarms. Earlier detection.

Cleveland Clinic: Pilot Performance Plus Platform Maturity

Cleveland Clinic evaluated both model performance at their pilot sites and the maturity of the Bayesian Health platform itself, including the research behind the model, how it had been deployed at other institutions, and what ongoing model maintenance looked like. They had done internal model-building work and understood firsthand the cost and complexity of what they were evaluating against.

Getting Clinicians to Actually Use It: Three Different Paths to Adoption

Model performance earns the decision. Clinician adoption earns the outcome. The panelists were refreshingly candid about how hard the latter is, and how different their approaches were.

MemorialCare: The Physician Champion Model — 90% ED Adoption

MemorialCare started implementation in the ED, which offered a structural advantage: a single, cohesive physician group with a long track record of participating in quality initiatives. They recruited dedicated physician champions — practicing clinicians with natural leadership instincts, not administrators — and built weekly adoption review meetings around their feedback.

The result: 90% adoption in the emergency department.

"We used the method of selecting physician champions to really be the people on the ground who would lead and champion the effort," Dr. Leo explained. "These are doctors who had not had other leadership positions, but clearly had the strong affinity and propensity for leadership."

Monthly deep dives into every "fallout case" — patients where a provider acknowledged the sepsis flag but didn't complete the expected clinical interventions — kept the program accountable without being punitive. The physician champion carried that feedback directly to the individual clinician.

That accountability structure shows up in the outcomes. When providers engage with the Bayesian flag, MemorialCare has seen a 3.6% absolute reduction in sepsis mortality — meaningful at an institution that had already driven mortality down 75% over more than a decade. Time to antibiotic administration dropped by 50%, and by approximately 66% when providers engage within the first hour. 

"Those are the kinds of numbers that make clinicians look twice and stop and pay attention," Dr. Leo said — and they were enough to earn system leadership approval for expansion to MemorialCare's remaining adult hospitals.

URMC: Non-Interruptive Technology Design — From Less Than 1% to 70% Adoption

URMC's ED providers were burned out from years of alert overload and explicitly asked not to be included in another alerting project. Rather than fight that, URMC redesigned the alert workflow entirely: Bayesian Health notifications begin as a passive sidebar indicator for the first hour, then surface as an Epic Haiku push notification to the provider's phone, formatted to look like a secure chat message rather than a clinical alarm.

The result: adoption went from less than 1% to approximately 70%, without making the alert itself more interruptive. "The fact that we're in the seventies is double our best alert," Dr. Gleber said. "Number two is probably somewhere in the thirties."

Sustained adoption was reinforced through a monthly personalized data email — each provider receives their own retrospective sepsis case summary, turning abstract quality data into something that resonates at the individual level. The approach also seeded ground-up quality improvement projects: individual units and service lines began requesting their own data and initiating their own improvement work, without being pushed from the top.

Cleveland Clinic: Enterprise Scale Across 20 Hospitals

Cleveland Clinic's adoption challenge was a different order of magnitude: going live system-wide across 20 hospitals with varied local cultures, clinical priorities, and starting points. Dr. Morrison's team found their highest adoption at sites that had been struggling most — a signal he read as validation, not coincidence. "It shows us that we're giving a new way to tackle this — that the teams are excited to have a new way to find septic patients, kind of break out of the burnout of some of the old metrics."

The infrastructure that made enterprise adoption stick: a multidisciplinary implementation team spanning clinical, IT, nursing, and quality; local ownership of the data; and an OKR-driven governance structure that kept sepsis mortality as the North Star metric regardless of campus. "What really matters isn't setting the goal at the top level," Dr. Morrison noted. "It's having the local team have the data and the will to engage it."

Five Lessons Worth Stealing

For health systems earlier in this journey, the panelists' experience points to a few principles that hold across institutions regardless of size or starting point.

1. Lead time and alert volume are the metrics that matter. All three institutions validated Bayesian Health primarily on whether it would deliver better lead time and dramatically fewer false alarms — not just whether the model outperformed on an AUROC curve. Clinician buy-in depends on a tool that cries wolf less, not just one that's technically superior. Get those two numbers right and the adoption conversation changes completely.

2. Start where you have structural advantage. MemorialCare started in the ED precisely because a single physician group meant faster alignment and faster wins. Know where your implementation tailwinds are strongest and build credibility there first before expanding.

3. Make the feedback personal. Both URMC (individual provider emails) and MemorialCare (personal fallout case reviews) found that personalized, specific data lands harder than aggregate dashboards. Clinicians respond to their patients, not system-wide averages.

4. Non-interruptive design is adoption design. The way an alert is delivered shapes whether it gets used. URMC's decision to launch with a passive sidebar before escalating to a push notification — designed to feel like a colleague's message rather than a system alarm — was the single biggest driver of their adoption turnaround.

5. Operationalization never ends. Weekly adoption meetings, monthly fallout reviews, unit-level data access: the institutions with the best outcomes treated go-live as the beginning, not the destination. The program maintenance is the program.

Watch the Full Webinar

This post covers the highlights, but the full conversation goes deeper — including Q&A on bundle compliance tracking, unit-specific alert configuration, and enterprise governance structures.

Watch the Recording

Bayesian Health's AI-powered sepsis detection platform is built for the complexity of real health systems, not controlled trials. If you're wrestling with alert fatigue, provider disengagement, or a mortality plateau that existing tools haven't moved, we'd like to talk.