Stanford Health Care Focuses on Fair, Useful, and Reliable AI Models

How AI governance is set up varies from health system to health system, and some academic medical centers are sharing best practices. During a Jan. 26 webinar hosted by Manatt Health, Christopher “Topher” Sharp, M.D., chief medical information officer at Stanford Health Care, outlined his health system’s governance approach, which includes a responsible AI life cycle and a focus on fair, useful, and reliable models.

Stanford Health Care is one of 25 health systems participating in the Manatt/AAMC Digital Health and AI Learning Collaborative, a peer learning forum for exploring best practices and practical strategies for integrating digital health and AI into everyday clinical care and operations.

Sharp is a practicing physician but in his role as CMIO he spends most of his time working to make sure that technology works for Stanford Health Care’s clinicians. “That's been a really interesting role, because it started as a an adoption leader, it evolved into an optimization leader and champion, and now it's really become much more of a strategic asset,” he said. “How we take these types of technologies and enable our clinicians is a part of our overall business and clinical strategy, and AI is certainly pushing deeply into that same frame of discussion.”

At Stanford Healthcare, the mission is to bring artificial intelligence into clinical use safely, ethically and cost-effectively. “We are excited for and proud of using AI in administrative use. We think it's important to use it in revenue cycle, it’s important in compliance use. It's even important in making sure that we change our beds on time and turn over our ORs promptly,” he said. “But ultimately, we want to get to the point where we've brought it to clinical use, which is very important to us.”

Sharp said creating the data infrastructure and interoperability between platforms is an imperative. “You can't have data science without having access to your data, so it becomes a terribly important component,” he said. “The governance and oversight is also just a ‘no regrets’ activity. We all know that the better we are able to align to our system strategy and needs, the more that flywheel is going to spin faster and faster.”

He said Stanford Health Care execs realized that to take full advantage of AI, they had to create new capabilities and grow new muscles. “That's where we identified the need to create more of a ‘center of enablement’ capability,” Sharp said. “For us, that meant recruiting some data scientists, putting leadership in place, and making sure we understood how that expertise is going to integrate into existing systems.”

Sharp said that Stanford Health Care’s chief information and digital officer, Michael Pfeffer, is fond of saying that they don't have a chief AI officer. “It's not one person's job to make AI work. At Stanford, we have a chief data scientist. It's one person's job to know what's good data science and what's not, but all of us participate in the question of how we're going to actually use AI to advance our organizational objectives,” he said. 

Lloyd Minor, M.D., dean of the School of Medicine, has launched what's called the Responsible AI for Safe and Equitable Health, or RAISE Health.  RAISE Health is a joint initiative between Stanford Medicine and the Stanford Institute for Human-Centered Artificial Intelligence (HAI) to guide the responsible use of AI across biomedical research, education, and patient care.

Sharp said this is a way of bringing the best and the brightest minds together to ask the tough questions around how to proceed. 

Speaking about the importance of governance, he noted that it is critical that they link to Stanford Health Care’s overall organizational strategy. “You need to have an executive-level sponsorship that can drive what is really the enacting layer that engages at the various levels below, making sure that we engage people and the workforce, making sure that we engage technologies and technologists in order to be able to bring all this to bear.”

Sharp said what he finds provocative in his organization, is that the C-suite leadership actually engages in the executive committees. “They don't defer or delegate that out so that it's done to report it back to them about how it works. They actually sit in those committees and spend the time with us, making sure that we understand where we're going, what we can do, and how we will actually execute and do this in our organization.”

He said that in the rubric of people, process and technology, you need processes in order to be able to manage this. Sharp described three key components they have developed. The first is a responsible AI life cycle. “There are endless products, endless solutions, and seemingly endless problems to be solved if you listen to the market today,” he said. “We really needed to make sure that we had a method responsible to our organization, to know that these items, as they come into our organization, whether they come in as a problem or a solution, will be funneled all the way through a process in order to make sure we can make the best decisions.” They use a rubric called Fair, Useful and Reliable Models (FURM) that was created by the data science team in the School of Medicine. 

The FURM approach allows Stanford Health Care to understand the problem-solution match, and then assess how they are going to approach that. 

Stanford Health Care also has developed a way to monitoring solutions, “which we've found to be critical, even as we begin to make sure that we create sustainable, valuable tools in our organization,” Sharp said.  One aspect of monitoring involves understanding the system and making sure that they can support the system integrity over time. The performance gets into the data science of how models actually work and how they monitor them over time. They also have operational impact metrics.

Chat EHR 

Sharp gave a concrete example of how they handle new developments in the AI world. One was when ChatGPT was released. 

“We didn't know how it would be used. That includes whether protected health information or other proprietary information would be exposed in that platform. So we went about creating a secure environment where we could allow for full experimentation by the entirety of the organization,” he said. They called it Secure GPT to help the workforce understand what's secure and what's not. They created it and began to watch its use. “In the spirit of a learning health system, we could see how it was being used, what it was being used for, and out of those use cases, we could derive what we should really focus on next,” he said. 

They chose to bring that data and information in a frictionless manner into an interactive, generative AI platform, which became a tool they built called Chat EHR. It offers the ability to interact with medical data by way of a chat as well as other interfaces. 

Sharp noted that Chat EHR looks at EHR data, but not only EHR data. It can look at other data as well. “You could start to feed multiple data sources in and then use multiple compute engines on the other side to pull insights out. We think that's an incredibly important asset, and something that requires a lot of architectural discussion about where your data sits, why it's important, and how you create more use cases into the future.”

Seeing common patterns in how people interacted with the platform led to the creation of automations. “We could find, for instance, activities that were being performed over and over in this chat interface, and ultimately realize we could codify those in a way that now they become an automation,” Sharp explained. “They could either be automatically triggered when a certain event happens, or at a standard interval to bring forward those data.”

He said this evolution of moving from a very big, broad, open platform to a platform that is really contextualized around patient information, then bringing that all the way to automations that really matter has been profound for Stanford Health Care. “Part of the challenge with AI is finding the problem and solution match, right? We have people who understand many problems in the organization, but don't understand how AI can help them, and we have people who understand how AI works, but not which problems are right to connect with. So this has been a tremendous learning evolution that we've been on.”

Thinking about ROI

Part of the new challenge with AI, he added, involves identifying the successful use cases and growing them and quickly identifying the unsuccessful use cases and killing them. Part of this is, he added, is around aligning against the key drivers that they care about and understanding the key problems to frame what the ROI should or could be as they bring in these different models, whether they are digital health models, AI models or combinations of those. “AI has the power, depending on where we put it, to really allow us to transform. If we focus on using AI to replace humans, we will miss out on the opportunity to get into places we could have never even imagined we could be when AI works alongside humans, and we think that that's a huge opportunity, and we want to invest in areas that will lead us into that in the future.”

It used to be the case that you could have a department say something looks interesting, let’s try it and see how it works. “Today, that really fails for two reasons,” Sharp explained. "One is it will die because it's not actually integrated into a larger strategy. By definition, that is going to be money sunk. The second is that we just have to think about the return on investment and the value proposition globally before we actually embark on this work. The question then becomes: Does your organization have a way to talk about investment that everybody can understand?”

Stanford Health Care has tried to divide that up into hard value/soft value questions. The hard value looks at a few key performance indicators that they care about. Sometimes those are direct revenue or savings, and some are things that are absolutely intrinsic to the survival of the organization —  things like length of stay, readmissions or where demand outstrips capacity substantially. “Anything that eases that burden actually becomes a return on investment for us and actually has a hard value,” Sharp said. 

On the other hand, there are soft values that can't be dismissed. “We use AI scribes, not because we see more patients, but because we know that our doctors actually see patients better and in a way that is better for them,” Sharp said. “I would encourage organizations to be able to do that prospectively. We do that as a part of that FURM assessment. When we're doing AI, we say, is it fair, useful, reliable, and part of that is does it bring value? How do we actually assure value and have that go through the governance to make sure that that's vetted before we get started?”