In Part 1 we explored the development of Duke Health’s Sepsis Watch™ and the capabilities built around the model that led to its success and durability.
Sepsis Watch™ has been in continuous operation at Duke hospitals for over seven years. It has increased Sep-1 bundle compliance, an important Medicare quality metric, by 89%. And while sepsis-related stays increased across the U.S. from 1.8 million to 2.5 million between 2016 and 2021, Duke University Hospital reduced mortality from sepsis by 29% during that time.
So, why hasn't the whole country benefited from a solution that works for clinicians and delivers results system-wide?
Inbound interest sparks a distribution question
Healthcare professionals at hospital systems outside of Duke learned about Sepsis Watch™ through a combination of high-impact academic publications, press coverage, presentations at conferences, and strategic institutional relationships. As the innovation gained traction, and the government implemented requirements around sepsis monitoring, Duke began receiving inbound interest from other leading health systems.
Because Sepsis Watch™ was one of the first deep learning integrations in a hospital setting that scaled system-wide, the team at the Duke Institute for Health Innovation (DIHI) focused on documenting how the AI solution fit into the messy reality of a hospital’s workflows. This grabbed additional attention from those who were interested in navigating the challenges of AI model and solution integration into clinical workflows and wanted to explore pathways to improve both quality of care and economics.
DIHI forged a relationship with the nonprofit research institute Data & Society to observe and interview nurses prior to and after the implementation of Sepsis Watch™. The research report by Data & Society’s Madeleine Elish generated interest and press coverage of DIHI’s work on early warnings for sepsis.
Another leading health system even came to do a site visit at Duke Health after Sepsis Watch™ had been running for less than a year to understand how it was being used on the frontlines of care.
Duke felt the solution could help other healthcare delivery organizations improve sepsis care. The DIHI team worked with the Office of Translation and Commercialization at Duke to explore pathways to get the technology to scale.
In 2018, this was a largely uncharted frontier for all academic medical centers (AMCs). While AMCs were accustomed to patenting molecules and medical devices, "selling" an AI/ML algorithm felt fundamentally different and presented several unique challenges that caused both DIHI and Duke to pause.
Technology transfer offices were optimized for intellectual property (IP) with clear physical boundaries. Software, specifically black-box AI, didn't fit the traditional mold. There were lingering questions about code vs model weights, responsibility for missed diagnosis, need for local validation, and more. In 2018, the industry lacked a standard playbook for how AI could be delivered to a customer. Commercial vendors were primarily focused on the model, while DIHI insisted that the model combined with the workflow was the actual product.
At that time, there were very few successful examples of AI point solutions commercialized by academic medical centers, so DIHI deliberately chose an approach designed to build trust before pushing ahead with broader commercialization. Rather than asking the market to believe in a promising model alone, the team first demonstrated that the Sepsis Watch model and workflow solution improved meaningful clinical outcomes at Duke, including reduced inpatient mortality among sepsis patients. In addition, DIHI showed that Sepsis Watch™ improved SEP-1 bundle compliance, providing further evidence that the solution could drive both quality and operational performance.
Before Vega Health, commercialization pathways for an effective AI model were limited to:
- Open source publish the code;
- Sign an exclusive agreement to license the model to a point solution company;
- Leave the institution and start a company around the innovation.
Duke took the second route. It granted exclusive licensing to a company with existing hospital products abroad looking to enter the U.S. market. The external company agreed to license three of DIHI’s models – sepsis detection, mortality, and cardiogenic shock — with the knowledge that implementing and monitoring a single model at any location would be too expensive to be worthwhile.
Scaling Beyond Duke: The Summa Health Experience
The company that licensed Sepsis Watch™ brought the sepsis detection model to Summa Health, a community health system in Ohio. Like hospitals around the country, sepsis prevention, detection, and treatment were clinical priorities for Summa.
The company and DIHI jointly applied for Small Business Innovation Research (SBIR) funding from the National Institutes of Health to enhance, validate, and scale Sepsis Watch™. The funding would extend and enhance Sepsis Watch™ to emergency departments, general inpatient wards, and intensive care unit (ICU) settings across multiple health systems in the United States.
The project had two phases and lasted three years. Phase I tested Sepsis Watch™ by analyzing historical data from two hospital systems to refine data collection and get clinician feedback. Phase II shifted to real-time operation, refining the AI model and interface with the ultimate goal of deploying the tool across any U.S. hospital.
The local validation at Summa Health proved that the model built using data from Duke Health performed well in a different clinical context. This was a novel finding and a major win for the DIHI team.
But the subsequent experience of integrating the model at Summa revealed that the change management required to use Sepsis Watch™ in practice at Summa Health was an even bigger challenge than anyone expected.
The company built a platform, user interface, and performance dashboards to surround the early sepsis identification model, but the learnings from Duke only translated up to a point given the vastly different clinical environments.
For example, the rapid response team (RRT) at Duke Health was extensively trained on how to incorporate the early identification of sepsis into their workflow. They were able to allocate several FTEs to the oversight of Sepsis Watch.
Summa Health had organizational barriers that prevented them from benefitting from an early warning system, including different practice standards and less FTEs dedicated to the use case. The early detection model was not something the organization could accommodate without hiring more staff.
It proved difficult to push for adoption of a model that staff perceived to add work to their days, rather than take it away. To help with the change management, the point solution company brought in Dustin Tart, the nurse manager who had led the Sepsis Watch™ implementation at Duke, to share how his team incorporated the early detection predictions into their workflow.
"It's one thing to hear that message from a startup or a technology team,” Suresh Balu, program director for DIHI, said. “But hearing it from the nurse who had to build that bridge between the solution and a spectrum of adopters — some more interested, some less — that's different."
Even as Tart’s visit increased adoption, the 2024 acquisition of Summa Health by venture capital firm General Catalyst shifted organizational priorities, and the sepsis prediction work fell off amidst the change.
The experience of commercializing Sepsis Watch™ exposed a core limitation of point solutions in clinical AI: a model without a clinical workflow, and dedicated implementation support, is a model without a future. The success of Sepsis Watch™ at Duke hinged on the deep collaboration between DIHI and frontline staff at Duke University Hospital.
This experience reinforced a core thesis that is now baked into Vega Health’s business: implementation is not a handoff. There is no such thing as a turnkey AI implementation in healthcare. It is a collaborative process that requires a team that understands both the technology and the clinical reality of the people being asked to change their behavior.
The Implementation Imperative: How Vega Health is building on the learnings from Sepsis Watch™
The commercialization challenges of point solutions directly informed the development of Vega Health, which recently licensed the early sepsis detection model from Duke and is working side-by-side with our customers to configure localized workflows that work for them.
Whereas the earlier commercialization partner offered a solid technical solution, it did not have the capacity to cultivate the relationships and trust among diverse stakeholders needed to effectively drive clinician education and change management. Vega Health’s platform, portfolio of solutions, and implementation approach allow us to invest in these processes.
Balu often says that, “Customers are not buying algorithms; they are buying outcomes.” Achieving those outcomes requires technical infrastructure, workflow integration, local validation, change management, and ongoing monitoring that Vega Health provides.
The hardest part of clinical AI is not the algorithm. It's everything required for the integration to be successful. This starts before any technology comes into play and before AI makes any predictions. There must be effective coordination and alignment across clinical roles, as well as strong clinical and operational champions. Information technology (IT) leadership also must have buy-in to make the best use of the technology infrastructure.
To achieve this alignment, Vega Health commits time to sitting down with frontline clinicians, administrators, and informatics leaders to understand their priorities and pain points. We visit the organization, we walk the floors, we commit time up front to pinpoint the issues alongside the people that are living them. It takes trust, which is built slowly.
What Health System Executives Should Know
For health systems evaluating a clinical AI investment, here’s what you can learn from innovators implementing AI for over a decade:
1. The solution is not the technology. It's the initiative. Sepsis Watch™ works best when it is embedded in a health-system-wide commitment to sepsis care, with buy-in from frontline nurses, nursing management, and clinical leadership. The technology amplifies an initiative; it cannot substitute for one.
2. Workflow localization is non-negotiable. The right alert, to the right person, at the right moment looks different in a community ED than in a tertiary academic medical center. Implementation requires discovery and context-specific configurations.
3. Peer credibility accelerates adoption. The most effective change management at new sites has come from clinical peers, not vendors. Connecting prospective customers with nurses and clinicians who have lived the implementation is among the highest-leverage activities in the onboarding and implementation process.
Interested in our work? Please contact info@vegahealth.com to learn more.
