In this second part of the discussion, Joe Kelly discusses the steps Mercy has taken to protect patient privacy, highlights partnerships that have kept them on the cutting edge and explains how Mercy’s foundresses act as a guiding light for his team as they move forward.
In this second part of the discussion, Joe Kelly discusses the steps Mercy has taken to protect patient privacy, highlights partnerships that have kept them on the cutting edge and explains how Mercy’s foundresses act as a guiding light for his team as they move forward.
To listen to the first part of this discussion, click here.
Brian Reardon (00:08):
Welcome to Health Calls, the podcast of the Catholic Health Association of the United States. I'm your host, Brian Reardon, and with me is Josh Matejka. He serves as executive producer of Health Calls. Josh, again, good to have you here where this is part two of our conversation with Joe Kelly who serves as Executive vice president and Chief Transformation Officer of Mercy. I'm just going to dive right in with Joe again. Joe, you hung around for the second part, so I guess the conversation's going well so far, so appreciate that. Let's, and I think the comments you made in the first part of this episode, again, I think we're just spot on as far as the role of, again, our mission, those who are working to advance our mission and their input on some of these changes that are constantly occurring with technology. For the second half of the conversation, I do want to get into again, more some of the practicalities and I think starting with artificial intelligence, we've been talking a lot about that during this season. It's obviously on the minds of everyone. It just seems like that tends to be a big focus, but I think it's interesting, artificial intelligence is not something that you consider to be new at Mercy. And maybe you could share a little bit about how long you've been working in that space and how that's really not a new development from your perspective.
Joe Kelly (01:27):
Yeah. First of all, we're blessed to have an incredible team in this space. About 11 years ago, we introduced our first artificial intelligence model. We have about 350 plus AI models running every day at Mercy. Everything from our robotic process, automations in the back office to machine learning models that are more in the clinical or operational space all the way up into generative AI for workflow and experience improvements. And I think back in episode one, I referenced kind of those guiding principles on AI that kind of shape our perspective on those and that council on AI and our mission team and our ethicists are involved in those decisions. We've also, mercy is a founding member of chai, the Coalition of Health ai, and we'll be the first health system in the country to actually go through and implement and install an AI model using CHAI's guidelines, which as a founding member, we help to help to create.
(02:29):
And I think that establishing data governance, having that theological reflection mindset of data governance and the role data plays in outcomes and in experiences. And I think lastly I'd say I, if you view AI as a point solution, it's probably not going to be as beneficial if you view AI as a force multiplier and an enabler that you can scale. And we often say you can't have AI without ia, without infrastructure and architecture. And I think so much of healthcare with disparate EHRs and nonintegrated systems, and we kind of live in point solution purgatory. And if those point solutions that we purchase are technology that we purchase software applications, if it's not integrated into the ecosystem, your AI efforts are probably not going to be able to scale. And if you don't have cloud computational capability, you're really not going to be able to benefit from scalable generative ai, orent ai.
(03:31):
And so I think being really intentional about the infrastructure investments, reducing of the tech debt and to be able to scale solutions across all of our ministry was really important to Mercy. Particularly for this podcast, it's particularly important to Mercy because we don't want a patient coming into a level one trauma center and getting different care models are different care than a rural patient going to a critical access hospital. That AI has to be available for all that we're privileged to serve. And that's where those principles, that's where our mission and ethicist, that's where those guidelines, that's where that theological reflection all come to hit home. It's that everybody was created in the image and likeness of God. We can't have disparate ways that we serve those in our communities.
Brian Reardon (04:18):
So 350 different AI models that you've got running at any one time. And this again, my maybe being a little naive here, but is there a master AI structure that sort of wraps all that together and has those, I mean, to me that's sort of mind boggling and how you get your arms around all of that and have that integration that's going to be possible. So I guess in layman's terms, how do you manage, again, especially we're going to talk about working with other partners that have their own AI solutions that they're bringing. You mentioned trauma care and how maybe an application running in a intensive care unit and treating a trauma patient, how does that interact with perhaps somebody in case management who then weeks later that patient is being discharged and maybe there's some AI solutions to finding out what their post-acute care is going to be. And so I guess how does that all get sorted out and managed and again in three minutes?
Joe Kelly (05:18):
Well, I think a few years ago when we centralized our data and analytics functions and we created Mercy's enterprise data and analytics office, that really enabled us to begin the process of, I mentioned the architecture and the infrastructure. We've really done an incredible job. Our Mercy technology services team, our IT team and our leaders in that space have done a fantastic job of eliminating tech debt, of integrating solutions and having enterprise standard software. Because of that, because of that standardization that we've undertaken several years ago, we now have visibility from Power BI or a data visualization perspective where we can see in near real time. And because we've moved to the Azure Cloud, we can see these things in near real time. So we have, we're on a journey. We have dashboards for everything. We need to whittle those down and have dashboards for the most meaningful things and actually move using AI as decision intelligence so that we don't need an operational leader to look at a dashboard.
(06:22):
We need AI to actually perform or recommend the next best action in real time. And that's the journey that we're on. But that enterprise data and analytics office and our IT teams, because we have standardization, because we have centralization, and because we've invested in infrastructure and architecture, it's very easy for us to see how those models are performing in real time. One of the most important things, especially in healthcare, is this, there's a term model drift, and it's very real. You can implement an AI model, a machine learning model. It could be trained on hundreds of thousands or millions of patients, but as it's used more and more, the model is learning, it's machine learning, and it could actually perform better or worse than you expect. And so we have a machine learning operations team that actually actively monitors all of these models for model drift and to ensure that the bias is minimized, the accuracy is optimized. And so the whole machine learning operations function is critically important, especially in healthcare because you can't have models that are monitoring the wellbeing of patients, not perform to the standard that you expect.
Brian Reardon (07:31):
Yeah, interesting. We've worked with a lot of partners, so I want to talk about, for example, the Microsoft partnership, and maybe a lot of our listeners don't know about that. So maybe that's a place to start. I know I've heard you talk about this in other settings, but really interesting as far as harnessing all of the data that's out there and really making decisions that are going to benefit public health all the way down to individual patient health. Can you just provide a little bit of overview about the Microsoft partner, I'm sorry, the Mayo Clinic partnership is what I meant to say. I know, I think you also work with Microsoft, but I was actually meaning the Mayo Clinic partnership that you're doing.
Joe Kelly (08:08):
Yeah, absolutely. And so it kind of goes back to that mindset that it's not our data, it's our patient's data, and it's our obligation and privilege to steward it. Our relationship with Mayo was really born of shared and common interests around getting to personalized, predictive, proactive care near real time through the appropriate channel. And that element of prediction using machine learning and building algorithms and models that can take the experience from millions of patients who have already been served and use that data to inform the outcomes and improve the outcomes of patients of today and tomorrow, I personally think is a beautiful reflection of our mission. And it ties back into the importance of data and the importance of how data can and should be used to inform patient outcomes. But that relationship with Mayo, so right now today, Mayo has access to all of our patient's data and we have access to all of Mayo's patient data.
(09:12):
Soon there'll be 12, 14 other health systems around the world where Mercy and Mayo will also have access to that data. So that sounds scary, but what we've done with that data is de-identified it, and we've paid a third party to come in and certify that we've not only de-identified structured data, the unstructured data and the notes, we've changed birth dates randomly. We've done everything, even imaging, I'm probably going to get the percentage wrong, but one of our physicians informed us that about 2020 1% of time the slices from a CT scan can be put together and then you can identify patient data through imaging. So we've spent literally years de-identifying data structured and unstructured. And that's important because it goes back to that mindset and that governance perspective and that mission orientation that it's our patient's data, not ours. So we could have easily joined another data collaboration where we would just send our patient data out and rely on them to do that work.
(10:21):
We did it ourselves, and I think it took longer and at greater expense, but our patient data doesn't leave our environment to go to Mayo. It's a data under glass strategy, meaning that physicians, researchers can look at our data, but our patient data remains de-identified and it never leaves our environment. So at Great Paine and expense, we built that because of that missional and foundational orientation. So why did we do that? Well, I mentioned before, when you're building machine learning models, bias can present itself and depending upon the geographic region of the country you're in, that bias can present itself in ways where one model can say this is a good way to treat a patient, but it may not be based upon their genetics or their ethnicity or their life experiences or their social determinants. So by having multiple health systems de-identify their data, build the same common federated environment to store that data, we now have more than 40 petabytes of data, of patient data where we can build machine learning
Brian Reardon (11:28):
Models. Put that in perspective. What was the bytes again?
Joe Kelly (11:30):
Yeah, sorry guys. Petabytes. So think of every novel ever written in every language since the beginning of time. That's about 38 petabytes. So we have a lot of data. Well, what do we do with this inheritance? As Catherine McCauley said when she founded out the Sisters of Mercy,
(11:49):
Well, we've chosen to take this data and use it for the good to build machine learning models to get to prediction, to identify disease states before they would occur or to prevent disease states from occurring. And so we're in the, I would say somewhere between the top and bottom. We're in the middle of the first inning in going here, but we have the capability today to build machine learning models with patient data from all over the world to reduce bias, to improve accuracy. And for every model that we build, we're going to put a nutrition label on it. And what that means is just like a box of cereal, our physicians, our clinicians, our doctors, are going to be able to see how this model was built, what mathematical technique was used to build it,
(12:33):
What the ethnic composition is of this model, what the gender breakdown is of this model, and then four different accuracy scores for this model. And so that gives our, again, our philosophically, we don't really view it as artificial intelligence. It's augmented intelligence. It enables our doctors to make decisions better than they otherwise would have. Between 1850 and 2020 and 2020 to 24, the same amount of medical data has been created. So the cognitive burden on our doctors is vast. This is a way for us to help use the experiences of past patients to inform the treatments of future patients.
Brian Reardon (13:11):
So in practical terms, an oncologist working at either Mayo or Mercy or any of these other systems that could join, could go in and say, Hey, I've got a patient with this is the diagnosis. It's maybe a more rare cancer. And they can really, by taking all of this information, potentially have a much more accurate a diagnosis and treatment plan based on all of the knowledge that's out there.
Joe Kelly (13:36):
That's correct. And they can see, if I'm a physician, I can see a patient with a likely diagnosis. I can then go in and see and identify other patients around the world who had the same diagnosis and look at all the various treatment plans and the outcomes of those. And I can consider genetic and gender and ethnic and regional variations and considerations, and that helps to inform my treatment plan while the data is de-identified. It's not just a data collaboration with Mayo. This also enables our physicians to have second opinions in real time. And so our goal in the coming years is to not only have an incredible data set from which we can better serve our patients, but to enable our physicians to have access to second opinions in real time rather than going back to dignity, rather than having a patient wait weeks or months to find out how to get into another system or another doctor or another. So again, it's putting that other centric mindset and that mission lens on the front end that this is why we've entered into the relationship with Mayo.
Brian Reardon (14:39):
And you think about the stress caused to patients in the waiting. And so that ability to get more instantaneous results or answers, I think that speaks to the incredible potential that this big data AI approach to medicine really offers.
Joe Kelly (14:55):
Yeah, completely agree. This is the issue and the opportunity of our time.
Brian Reardon (14:59):
Other partnerships that you would just want to highlight, and I apologize to Mayo for confusing them with Microsoft, but it seems like you mentioned cloud computing. I know a lot of folks are on Microsoft. Obviously there's a lot of tech companies out there. And I guess the reason for asking that is, again, as you're trying to bring their data set or their technology approaches to what you've built, can you talk us through how you work through those type of partnerships?
Joe Kelly (15:22):
Yeah, and I did the same thing, Brian. I referred to Microsoft as well, and ironically, I think it was either tremendous foresight or a Freudian slip, but Microsoft is one of the partners I'd love to highlight. And I think to Microsoft Credit into Mercy and our team's credit, they viewed us as a customer. We viewed them as a vendor. I think we both view each other as a partner now, and that journey the past few years has been tremendously beneficial. We've signed a 10 year relationship with Mayo. We've signed a five year strategic partnership with Microsoft, and Microsoft I think recognizes that finding a partner in healthcare with the intentional infrastructure and the integrated architecture we've built for scalability is rare. And as a result, they've invested their time engineering resources and money in us. They've helped to fund our building, our intelligent data platform.
(16:19):
That's what we refer to as our Azure based cloud computational platform that all of our AI models run and scale off of. We are one of the heaviest utilizers in the country of their technology centers. We have hackathons there on a regular basis, and we've developed some proprietary solutions like patient handoffs at these hackathons that have dramatically improved patient outcomes as well as reduced administrative burden for our overburdened nurses. So the journey with Microsoft has been important. And going back to the, you can't have AI without infrastructure and architecture, without ia, there's, there's no shortcuts. You got to take years to actually, especially in healthcare, with so much disparity and such little standardization and such high variation, you have to have a lot of intentionality. And I think our partnership with Microsoft has enabled the infrastructure and architecture for us, our partnership with Mayo on the AI front to bear fruit.
Brian Reardon (17:21):
And I guess to wrap up the conversation, collaboration, very key to what we do in healthcare. I think the more successful we are is when we build those really authentic trusting relationships where goals are aligned, where missions are aligned. So to kind of wrap things up, Joe, can you talk a little bit about maybe some tips or thoughts you would give to those listening that are looking at a partnership maybe with another healthcare provider down the street and looking to collaborate? Maybe it's with a technology partner. What are some of the sort of core principles that you would share as they enter into these partnerships?
Joe Kelly (17:58):
Yeah, great question again, so the first one that comes to mind is having a platform mindset. And what I mean by that is we need to create an ecosystem where us as mercy or other systems create an ecosystem to help navigate patients effectively. With 70,000 ICD 10 codes and 7,000 DRGs, we can't go it alone. We need partners, we need help, we need to, and if we put that mission lens on the dignity lens of really how do we help patients navigate and get the care that they need through the channel of their choice at the right time, we have to create an ecosystem and we have to enable that ecosystem through technology. Two, resist the temptation and the tyranny of a point solution. We found ourselves in point solution purgatory, and we've decommissioned well north of a thousand of applications that we thought were a good idea at the time. A point solution often creates more trap doors than they close if they're not integrated into your technology ecosystem. Third, data neutralizes emotion, lead with numbers, demonstrate results, credibility builds over time, not overnight. A lot of these models, I've referenced our generative AI nursing handoff that saved our nurses 250,000 hours annually, that also our physician ambient technology that helped free up 66,000 additional annual appointment slots for patients. And when we go back to what matters most to patients, you can have the best care in the world, but if you can't get in,
(19:38):
It doesn't matter. And so just really leading with numbers. Then the last thing I'd say is the sisters are pretty amazing. They gave us a pretty impressive North Star 200 years ago. We just need to follow it. While the times have changed and the tools have changed, and the enabling functions have changed, if we stay true to that North Star, if we think of it from a dignity, from a mission orientation, it's not easy, but it becomes easier for us to figure out how to deploy all of the tools at our hands or that we have access to better serve that mission and to better serve the people that we're privileged to.
Brian Reardon (20:14):
Great stuff. Thanks for bringing so generous with your time and sitting down with Josh and me and really talking through a lot of the aspects of this. I think it was a really enlightening conversation and I really salute the work you're doing. And again, thanks for spending time with us.
Joe Kelly (20:28):
It's a privilege to be here. Thank you guys.
Brian Reardon (20:30):
Again, that was Joe Kelly. He's executive Vice President and Chief Transformation Officer for Mercy. This has been another episode of Health Calls, the podcast of the Catholic Health Association of the United States. I'm your host, Brian Reardon, and joining me is our executive producer, Josh Matejka. We've had additional support for these episodes from Yvonne Stroder, this episode and the one before. It was engineered by Brian Hartmann at Clayton Studios here in St. Louis. Of course, you can find health calls on all of your favorite podcast apps and services, as well as on our website, which is chausa.org/podcast. If you enjoy the show, please give us a five-star rating, write a review. We'd love to hear back from you. Thanks for listening.