A.I. in Health Education with Richard Van Eck

Richard Van Eck, Ph.D., is the David and Lola Rognlie Monson Endowed Chair for Medical Education at UND’s School of Medicine & Health Sciences. In his role as chair of the School's Artificial Intelligence Committee (AIC), he spoke with the School's Dean Dr. Marjorie Jenkins about the transformational changes that artificial intelligence and machine learning have already produced in health care and health care education, and the challenge of integrating AI as it continues to evolve.

Transcript

SPEAKER_02 0:07

Welcome to Health in Action, a podcast from the UND School of Medicine and Health Sciences. In each episode, Dean Marjorie Jenkins chats with UND faculty and staff about the latest developments in healthcare education, our changing healthcare landscape, and the latest news from North Dakota's only interprofessional school of medicine and health sciences. Richard Van Eck is the David and Lola Ronley Monson endowed chair for medical education at UND School of Medicine and Health Sciences. In his role as chair of the school's artificial intelligence committee, he spoke with Dr. Jenkins about the transformational changes that artificial intelligence and machine learning have already produced in healthcare and healthcare education, and the challenge of integrating AI as it continues to evolve.

SPEAKER_01 1:06

Well, good morning, Dr. Van Eck. May I call you Rick for the purposes of the podcast? So tell me, as you're thinking about meeting the today's learner where they are, have you had any aha moments as you have been doing a deep dive into artificial intelligence, for example?

SPEAKER_00 1:28

I would say that all of my readings and uh experience with AI is one long aha moment, in punctuated by brief periods where I think I've got my arms around it. And then I learn something new and think, oh, wait, that's a whole different area. So I think that's part of the challenge of teaching and learning and AI is being comfortable with the pace of continual change. There's some core precepts that we can focus on that are critical. For example, managing risk, understanding bias, um, data infrastructure, and uh evidence-based practices and the technology that we need to implement. Those are things that are going to continually be true for all AI, but the means of delivering AI changes. And I think there's a lot of things that we take for granted that turn out not to be so true. And I'll give you an example. You asked about learners. So one thing I learned from my colleague Anna Kinney on the AI collective at UND was that many students are now doing what we're calling human signaling in their work for their courses. And what they're doing is they're introducing errors that they think an AI would have caught because they want to make sure that their teacher doesn't think they used AI. In other words, they'll look at this and say, well, if you'd used AI, it would have caught this. So this must clearly be your own work. That's the thinking of the students, right? There's also a perception that students are using AI and are fluent with it. And it's a moving target. Some students are familiar with it, some students are using it, but there's a lot of research that shows most of them are afraid of it because they're afraid it's going to get them in trouble. They know that there's bias and risk involved. But at the same time, they'll use it for things like uploading their resume into an AI tool to get improvement without understanding that they've just given their address and phone number potentially to the entire world because they're using non-secure AI. So it's a huge mixed bag. One thing I think we can take away from this is we cannot count on students to teach themselves AI. They are not going to do a good job of that. We have to lead in that space.

SPEAKER_01 3:52

I could not agree with you more. And you mentioned bias, and you know, my background is in women's health and physiologic chromosomal differences between males and females, and how that portends future uh illness or health or prognosis or therapeutic response. So tell me a little bit about because it's almost like we go into chat GPT and it's magic, right? You answer, you ask some question, and then suddenly this whole plan will come out in front of your eyes. And so, in relation to health and the health of the people, what are some really um important things we should be thinking of about where that AI tool is getting that information? And what should we be thinking about and what lens should we be thinking about?

SPEAKER_00 4:47

Yeah, boy, that's a great question. And I love having these conversations because I think this is where both the promise and the peril of AI lie. So I don't have to tell you that the data on which a lot of our evidence-based practices and clinical trials are based is largely based on males. And so there's a huge gap in terms of whether these treatments and diagnoses and so forth are as appropriate for women as they are for men. Well, that is something that has moved forward into the age of AI because AI can be either unsupervised or supervised. In healthcare training, we want supervised AI. In other words, physicians and radiologists and experts are sticky are looking at the data, considering it, trying to mitigate bias, et cetera. But nothing can avoid the fact that the data that we're going to be feeding into AI comes from our healthcare system. And our healthcare system has bias in its data because of the structural kinds of things that exist. So, for example, there was a uh a company that tried to build an AI tool that would predict who would be the sickest in a hospital admission so that they could use it for bed planning, elective surgeries, uh, who would need to go to ICU, those kinds of things. And they used uh the IDC codes because they figure the more billing codes you get, the sicker you were. And the AI tool almost immediately said it was 95% accurate. And it turned out that it said white people are going to be the sickest, and black people will not be the sickest. And they thought, what is going on with this data? Well, the billing codes are associated with insurance levels, which are associated with economics. The data itself is biased because of the healthcare system. And so it was an honest mistake. Nobody thought they thought IDC codes are the way to go, but it turns out not. And this is just one example. Any little variation in the data that's there, even if it's one error, 1% error, can turn into almost guaranteed errors later on in the AI model. So we, it's we're never going to be able to make it perfect, but we have to be hyper-vigilant.

SPEAKER_01 7:03

Thank you for that great example. I think it's healthcare systems and our healthcare delivery systems are built on evidence and research that historically has enrolled more men than women, but even when we enroll more women, we all we don't disaggregate the data, right? We don't look at the data by uh sex or gender. And so we inevitably dilute the pool and we don't find a difference, or we find a difference that may not apply to both. When, as you're thinking about this program of building for our learners, what are your thoughts about humanities and AI? What does that mean to us in higher education and how what what sort of responsibility do we carry forward in this domain?

SPEAKER_00 7:54

Yeah, that's a great question. So I can answer it in a couple of ways, but the first way is that one precept of AI in any any usage, but especially in healthcare and in education, is that a human is always in the center of the loop. AI is not the decider. No matter what you've done, no matter how well you've trained AI, you cannot put AI in charge of it. There has to be a human who looks at it. And uh, working with Devin Olson, one of our library resources, and Sarah Westall, they shared a decision diagram one time that says, if you're gonna use AI, do you have the expertise to evaluate its output for accuracy? Number one. Number two, do you have the time to use that expertise to evaluate it? And then number three, are you willing to take full responsibility for any errors the AI makes that you did not catch? And that really illustrates for me that precept. Humans are in charge of the decisions that AI makes. Nobody should be letting the AI move itself. The other thing about humanities and human-centered computing are things like any use of AI incurs some level of risk, environmental risk, electricity cost risks, water usage risks, bias, uh, bad information, et cetera. So we have to be intentional about which things we use it for. For example, a lot of people will say, gee, AI is great for efficiency. Let's use it to do as many things as possible. But why would we pursue some of those things if, for example, in Hollywood animation, it's going to put an entire industry out of work? Is that where we should focus our attention, or should we focus is on deep learning for earlier prediction of cancer, right? Those are two different questions. And so there's a human and an ethical side of doing this and thinking about what the role of humans are with AI and the humanities as well, because we could use AI for a lot of things, storytelling, movie making, we're hearing about all of these things. Um, but are we losing something essentially human if AI is telling the stories and if it's building from the data that already exists? Is it just rehashing ideas? Are we ever going to get true innovation? I think these are questions we have to continually ask ourselves.

SPEAKER_01 10:17

Well, I think that is a great statement to wrap up on and more to come. And I hope you'll join me in the future as you're building out AI and uh and our AI hub. And there's so much more we could talk about today. But I think you have given us a lot of food for thought, and I really appreciate you being here and participating in our very first podcast. So thank you. Thank you, Rick.

SPEAKER_00 10:45

Thank you for having me. I enjoy the conversation as always.

SPEAKER_02 10:57

Founded in 1905, the UND School of Medicine and Health Sciences is the only MD granting institution between Minnesota and Washington State. In addition to its four-year program in medicine, the school houses degree programs in athletic training, medical laboratory science, occupational therapy, physical therapy, physician-assisted studies, and public health. It also hosts master and doctoral programs in biomedical science, clinical and translational science, and indigenous health. Since 1973, our historic Indians into medicine program has produced hundreds of indigenous physicians, therapists, lab scientists, and other health professionals for practice in rural and underserved areas. Learn more at med.und.edu.