In this episode, Chief Innovation Officer Jesse Goodwin, Ph.D., talks with Arman Kilic, M.D., Director of the Harvey and Marcia Schiller Surgical Innovation Center at MUSC. Dr. Kilic discusses how artificial intelligence is being used to advance medicine and treatment in transplantation, through machine learning. Through AI, new risk models are being created to improve the cross between the donor pool and patients awaiting transplants.

00:00 The start of the show

02:32 Arman Kilic joins the show

04:46 Leveraging AI in heart transplantation

10:13 Using machine learning to train models

13:35 Regulatory issues regarding AI today in medicine

19:30 Advice to students who want to learn AI, ChatGPT, and machine learning

This show is a production of the MUSC Office of Innovation and the Office of Communications and Marketing. Learn more about innovation at the Medical University of South Carolina (MUSC) by visiting: https://web.musc.edu/innovation

Arman Kilic

How do you take technology and algorithms and A.I., and how do we maximize the odds that it's actually going to be used in clinical practice? So as we build these systems, both the risk models and the dashboard, but also the A.I. framework, we're going to be in constant engagement with end users, and that would be cardiologists that would be heart surgeons, it would be nurses, it would be policy makers, it would be governing bodies, everybody to make this a very iterative process.

Arman Kilic

And in doing that, we maximize the odds that this will actually be a clinically useful tool.

Erin Spain

Welcome to the Innovatively Speaking Podcast, a podcast brought to you by the Medical University of South Carolina And each episode we dove into the origins of the next big things the who, the why, and how we explore ideas that are changing what's possible Here at the Medical University of South Carolina and in some cases across the world, my name is Aaron, Spain.

Erin Spain

I'm here in the MUSC podcast studio with my co-host, the Chief Innovation Officer here at MUSC, Dr. Jesse Goodwin.

Jesse Goodwin

Hey, Erin, thanks.

Erin Spain

I'm so excited for this episode. It's about artificial intelligence and machine learning. And this is a topic that has just exploded in the past year, and I'm really interested to hear how this is being used at MUSC.

Jesse Goodwin

Yeah, as I laughingly tell my team, it's an every day, all day lately. And today's guest is really fascinating. And I think we're going to learn a lot by talking to him.

Erin Spain

Yes today we welcome Dr. Arman Kilic. He's the surgical director of the Heart Failure at Heart Transplant Program and the director of the Harvey and Marcia Schiller Surgical Innovation Center at MUSC. He's at the forefront of revolutionizing the heart transplant process through the integration of artificial intelligence and machine learning. And his interest in AI and machine learning goes way beyond transplantation.

Erin Spain

So we're really excited to hear from him today.

Jesse Goodwin

Yeah, I got to meet Arman right after he came to MUSC so a few years ago, and I've gotten to know him better since MUSC received the gift from the Schiller family that allowed him to set up this center within the Department of Surgery. And he's been a great colleague to work with. He's got lots of big ideas, and he's built a strong team around him.

Jesse Goodwin

And I'm excited to learn more from him today about what he's doing in this space and about I just in general.

Erin Spain

Awesome. Well, let's bring on Dr. Kilic. Welcome to the show. Dr. Kilic.

Arman Kilic

Thank you for having me.

Erin Spain

Can you tell us a little bit about your interest in innovation, artificial intelligence and machine learning and where all of this stems from?

Arman Kilic

Arman Kilic

So I've been involved in analyzing and developing new risk models for that database for about 20 years or so at this point and working really at a national level and how do we better do risk modeling and how do we better account and use this wealth of data that we have both for patient outcome improvement, quality improvement, but also from a research standpoint in terms of research and development and looking at novel ways of managing and looking at data.

Arman Kilic

About ten to 15 years ago I started during my time in Pittsburgh, I found myself interacting and collaborating a lot with faculty at Carnegie Mellon University, which is a leading center in A.I. And we utilized a lot of the big data and health system data that was available there and started exploring different ways to model the data to look at data.

Arman Kilic

Arman Kilic

But I think how we can use AI and starting to think about some of the deficiencies we have in health care and health care delivery, that is really skyrocketed, I would say really in the last several months if maybe at the most last couple of years, it's starting to really garner a lot more interest in health care and how do we learn to apply in health care and so, you know, I feel fortunate because I have this background in working with data and in a I have been thinking about this for a long time and a lot of my personal work has been in this.

Arman Kilic

And now to see the field sort of taking off, I think, very well positioned to help move this field forward.

Jesse Goodwin

Let me ask you a little bit about your specific interest in heart transplantation. Can you talk a little bit first about the challenges and opportunities that you see within the field? And then we'll talk about how you're going to leverage A.I. for them?

Arman Kilic

To me, there's nothing cooler than to take what is a catastrophic event for another family, but give them an opportunity to help somebody else out and take that organ and transplant and somebody who's struggling with heart failure and is going to die without a transplant, and then to have them flourish and lead a long, healthy life with high quality of life.

Arman Kilic

It is a highly complex operation and a highly complex program that requires an extensive amount of care, coordination, workup, patient selection, perioperative and post-operative care. So that, I think is very appealing to me. It's a challenging field, but it's a very rewarding field. And then from a academic and sort of zooming out to more of a national scene, the other appeal of heart transplantation is that donor organs, because they're scarce, we have to be very smart about how we allocate donor organs.

Arman Kilic

Arman Kilic

And when you step back and think about it, managing that entire system involves a lot of data. There's a lot of donor data there's a lot of data on candidates who are on the waitlist waiting for a transplant. And then there's a lot of data once you've matched that donor with that candidate of how they do after surgery.

Arman Kilic

So again, with my data and technical head on, that appeals to me as well because it's an opportunity to do things better and to work with the vast amounts of data that exist in the transplant field.

Erin Spain

So you were recently awarded an almost $2 million R1 grant from the NIH to improve heart transplant demand and these organ supply issues with AI and machine learning. Tell us about this award and what you plan to contribute with this project.

Arman Kilic

Yes. So we're looking to disrupt how heart transplantation is done in the United States. Some of the challenges currently and some of the obstacles in the way heart transplantation is handled at a national level as one as these donor offers are being made, the centers, the decision making of whether to accept or reject a donor is highly variable.

Arman Kilic

It's very subjective, and it's not very data driven as an example, when we get a donor offer now for one of our patients who's on a waitlist, we hold a conference call. You know, the cardiologists, the heart surgeons, we all get together, have a conversation. If we think it's the right donor for that specific patient or not. But it's a very subjective conversation.

Arman Kilic

So one of the things we're aiming to do with the grant is take the wealth of old data on all heart transplant patients that have been performed in the United States. And we're looking to create risk models that utilize what's called deep reinforcement learning, which means you're taking into account the dynamic changes that are going to happen with the donor as well as with a waitlisted candidate.

Arman Kilic

So at point A, they may have a certain lab value, but that changes to point B when the donor offer is made and so forth. So the modeling approach we're taking takes into account all these changes that happen throughout time. And then at that cross section of time when the donor offer is being made to a potential candidate, we're able to use all that data that's been collected for both the donor and the waitlisted candidate.

Arman Kilic

And then provide a dashboard to clinicians that will help them in making the decision by showing them what is the projected outcome if you reject this donor. So if you reject the donor you're risking more time that that waitlist candidate is waiting for something for another organ. And so it's telling you what is the risk that that waitlisted candidate is going to die before they get another good organ or a better organ.

Arman Kilic

It also tells you what is the projected outcome if you accept that donor, because if you accept that donor and you end up transplanting that donor organ into the recipient, it will project and tell you what is the anticipated survival of that patient. So that's one arm of the grant. The second arm is one of my collaborators who's out of Carnegie Mellon, had actually developed an algorithm that runs the incompatible kidney exchange program for the United States.

Arman Kilic

And that algorithm is driven by artificial intelligence. And what we're looking to do is actually take that framework, the novel framework, and adapt it to heart transplantation. And the beauty of this framework is it's actually able to use the vast amounts of data, historical data, and it's able to project with very high accuracy what the future state will be, if you will.

Arman Kilic

So if you think about it, it's almost this dynamic marketplace that exists where you have donors, you have waitlisted candidates, and they're in flux constantly. New donors are being added into the pool. Donors are being taken away, new waitlisted candidates are being added, waitlist candidates, some are dying, some are being transplanted, and therefore being taken out of the pool.

Arman Kilic

And this API framework is able to not just take a very narrow view of what's happening, which is how the current system runs, but it's able to actually model everything that's happening in that entire system at once, but also project what's going to happen in the future to optimize it. So you're optimizing what donors get offered to which waitlisted candidates.

Arman Kilic

And then when they're offered, we're looking to optimize the decision making process. And our hope is by combining both of those elements that we're able to improve, number one, the number of transplants and number two, the outcomes of those transplants.

Jesse Goodwin

So knowing that the fundamental need for any artificial intelligence or machine learning model is having a tremendous amount of data in order to train your models on how many sites are you collaborating with in order to amass enough data to make those predictions.

Arman Kilic

So this is one of the huge advantages of transplantation is that the United Network for Organ Sharing which has a contract from the OPTN, which runs transplantation in the United States, it's a federal contract and it provides all the data and it's a very, very rich clinical database. So instead of having to go and individually collaborate with a lot of centers, we've actually partnered and Yunus, which is the United Network for Organ Sharing, is one of our collaborators so we are doing this with partnership and support from the national governing bodies for transplantation.

Arman Kilic

And what they've agreed to is to give us the customized data set that contains every single data element that's been input it for donors and for recipients. So this is going to be a massive, massive pool of data that were then able to leverage and it should encapsulate every single center that performs heart transplantation in the United States.

Erin Spain

You were talking about this subjective approach that happens right now where it's kind of people huddled. You're on a conference call you're talking about the candidate. Is this going to be a good match? What kind of burden is that on a surgeon? And how could this modeling and this dashboard, this project, help relieve some of that burden?

Arman Kilic

It's a huge burden. And one aspect I didn't mention is there is a time sensitivity to it. So typically you're given about an hour to make these decisions. Now, sometimes we need additional studies and, you know, they can get those studies and that can drive the decision making process a little bit longer. But typically they give you an hour because you can imagine these coordinated hours who are doing these phone calls to say you have this donor, it's available for this candidate that you have on your list.

Arman Kilic

You want to accept or reject. You know, they may have a list of hundreds of centers they need to get through. You know, if people are rejecting the donor, they have to keep trying to go because they're also incentivized to place these organs and to have these organs utilized for transplant and this objectiveness and the burden on the clinicians is very high because these are life or death decisions.

Arman Kilic

You know, if we make a suboptimal decision, not only are we potentially harming the waitlisted candidate, but we're taking a donor organ that could have potentially been used elsewhere and had a better outcome. So all these things really interplay. And that's why our thought is that by utilizing data, it may really help the clinicians, you know, in a fraction of time be able to pull everything and say here are sort of the projected outcomes of rejecting or accepting this donor.

Arman Kilic

And that may be helpful for small programs, for big programs, for experts, for non-experts. It's just data that we think will be very functional. And the part of the grant I didn't mention is we're actually also using implementation science experts who their focus is how do you take technology and algorithms and AI and how do we maximize the odds that it's actually going to be used in clinical practice?

Arman Kilic

So as we build these systems, both the risk models and the dashboard, but also the AI framework, we're going to be in constant engagement with end users, and that would be cardiologists, that would be heart surgeons, it would be nurses, it would be policymakers, it would be governing bodies, everybody so to make this a very iterative process, to say, here's a prototype, what's good and what's bad about this, and then we go back and we change the things that we need to change.

Arman Kilic

And in doing that, we maximize the odds that this will actually be a clinically useful tool.

Erin Spain

There are a lot of issues when it comes to regulatory questions. And I can you talk a little bit to that and what's happening specifically in health care when it comes to regulation?

Arman Kilic

Yeah. So there are a couple issues that related to that. One is trust the concept of trust by patients and by physicians in AI. So very interestingly, you know, if you take a look right now, there is this concern of this deteriorating patient physician relationship chip. And the question is, how would I fit into that? Would it further that gap?

Arman Kilic

Would it bring it closer together? There was a very interesting study where medical questions were posted into a social media post and a chat bot, which was an AI chat bot, responded to those questions and then physicians responded to the questions as well. And then the patient who were posting the medical questions scored in a blinded way, scored the chat bot and scored the physicians both in terms of the quality of response and the empathy.

Arman Kilic

And it is remarkable. I highly encourage people to read this paper. I believe it was published in one of the JAMA affiliate journals, but the chat bot was almost two fold higher on a scale of one to five on average, both in terms of quality of response and empathy. But the question becomes how does that impact the patient physician relationship?

Arman Kilic

And when you look at surveys, so there's been a lot of surveys done of patients and the general population to say, you know, various questions about A.I. and how would fit into health care and the area that ranks last but our patients believe this is going to worsen the patient physician relationship. And so we have to figure all that out.

Arman Kilic

That's quite a challenge. But understanding how it could be used as assistive technology, I think is very important. And how can physicians use that? And that's where a lot of the education piece comes into play as well. The other issue was trust is defining what is the threshold of risk or errors that would be willing to accommodate for AI.

Arman Kilic

So a very good analogy to this is self-driving cars. So if you take a look, let's say at a population level, self-driving cars may inflict less, let's say car versus pedestrian fatalities compared to human drivers. But imagine the media attention and sort of the backlash that happens when a self-driving car, you know, hits a pedestrian. But if you look at it statistically, it's lower.

Arman Kilic

There's a whole different perception of that. And the same thing may happen in health care, where if you have a fully automated A.I. system that makes a fatal error, even though that error rate maybe statistically less than physicians the question becomes, how is that going to be viewed by the public from a regulatory standpoint? The European Union, for example, just put out a sort of air act that looks at different levels of AI and assigns them risk.

Arman Kilic

And then there's sort of an action item that's tied to that. So things, for example, like prohibitive risk would be things like mass surveillance or, you know, using something for harm at a large scale. And those are prohibited uses of AI and then other things like, you know, video impersonations or these deepfakes there has to be a transparency requirement there.

Arman Kilic

So they've done a very nice job, I think, of assigning these various risks on AI tools and what the governing ruling is regarding that. There are things in the United States that are happening as well. There's a federal air act that's being discussed by Congress. The White House did release sort of a Bill of Rights. So these governing bodies are getting together with leaders and tech leaders in health care and so forth that are helping to develop some of these things.

Arman Kilic

But it's interesting, there's a map that shows you how many states have no AI regulation or no AI laws, and the vast majority fit into that category. So this is an area for all the lawyers out there. If you want to make a huge career air and in health care and figuring out the legal and regulatory aspects of that is a huge gray zone right now.

Arman Kilic

And it's going to need a lot of help and a lot of effort and figuring out how do we govern this space.

Jesse Goodwin

Switching gears just a little bit, you are the director of the Harvey and Marshall Schiller Surgical Innovation Center, which was established through a gift from that family and you are home to many, many projects outside of just your own. Can you give us a flavor of the types of projects that you're working on within your center?

Arman Kilic

Yeah, there's a lot of ongoing projects some of them I personally oversee that are in the cardiovascular space, but there are others, and this is really a department level innovation center. So we have projects in vascular surgery we have projects in surgical oncology, we have projects in trauma and pediatrics. And there's a lot of exciting new projects that are coming down the pike that we're supporting.

Arman Kilic

And one of the nice things about our center is that we are staffed with several Ph.D. or master's level scientists. We also have those that are experts in data extraction and working within the data environment and within Epic or other hours. And at the same time, we have surgeons that are partnered or involved with the center. The concept of co-location of both the clinical and then the data science piece we think is what accelerates a lot of these projects forward.

Arman Kilic

But a lot of our projects have to do with predictive analytics. So how do we use data regardless of what field or domain it's in? To better predict something, we have several projects that are imaging based. So you can think about carotid ultrasounds, echocardiograms, CT scans for cancer that we're using to map to various outcomes to see if we can build automated pipelines to detect various things that can be seen in imaging.

Arman Kilic

We're also leveraging large language models, which is sort of the newest iteration of A.I. that's gathering a lot of attention. And because of a lot of the investment we've had into hardware and server power, we're able to build some pretty powerful models. And it's a really interesting time because these large language models are interactive and things like chat bots or things where there are prompts like, you know, chat, but you can tune these to learn on medical data and you can imagine figuring out the path to how we use these large language models to interact with patients, but also to interact with clinicians is a huge challenge and a huge opportunity.

Arman Kilic

So we're doing a lot of projects to evaluate that interface between large language models and health care.

Jesse Goodwin

That's really exciting.

Erin Spain

What advice do you have for a listener, maybe a clinician or student who wants to better understand the AI machine? Learning space? How do they get started?

Arman Kilic

I think that's a great question. And the good news is there is a wealth of information out there. One of the resources I point a lot of people to is Coursera. A lot of people are familiar with Coursera. It's not just A.I., but just Education Archive, if you will, of a lot of different resources for different domains. And they actually have really, really great courses on AI.

Arman Kilic

Many of them are free, and many of them are offered by leading centers in AI. I think the other one is you sort of dove in, especially if you're on the health care side. You sort of start diving into projects and working with real problems. You can learn a lot from doing that. So if you are a physician, for example, if you just look up, whether it's A.I., it's machine learning, it's informatics degree programs that your university may have, you know, you can find some contacts there, start discussing with them that you want to learn a little bit more about this.

Arman Kilic

I would tell you that on the AI and informatics side, those departments within universities are begging to collaborate with health care providers because they love getting access to health care data, working on real problems, and putting their expertize to use to see how they can work on real problems and help solve those problems.

Erin Spain

Thank you so much, Dr. Kilic for being on the show. We appreciate your time and your expertize.

Arman Kilic

Well, thank you for having me.

Erin Spain

You've been listening to the Innovatively Speaking Podcast with Medical University of South Carolina. If you enjoyed this episode and would like to support the show, leave a reading and a review to hear more innovative ideas and to share your own. Subscribe to the show or visit us on our Web site Web MUSC Edu Slash Innovation.