A.I. and its potential to transform healthcare

View  Transcript

How could A.I. potentially revolutionize healthcare delivery? The application of computer algorithms to medical knowledge has a long history, one that has accelerated in recent years to include generative AI platforms like ChatGPT. U-M expert Dr. Cornelius James discusses how AI is touching everything from doctors’ workloads to diagnostics to medical education.

More Resources

Learn more about A.I. in healthcare

Learn more about The Fundamentals

Transcript

Kelly Malcom:

Welcome to the Fundamentals, a podcast where we explore biomedical research here at Michigan Medicine. Research is fundamental to University of Michigan's mission to improve the world. On each episode, we'll meet the people behind the research, learn more about their fields and the fundamental questions they are trying to answer. I'm Kelly Malcolm, a science writer and communication strategist for the University of Michigan Medical School. This season, we'll start by explaining a little bit of the history behind the questions our experts are asking and get a glimpse into the future of healthcare. In the summer of 1956, during a conference at Dartmouth College, a small group of scientists gathered to explore one idea.

That "every aspect of learning or any other feature of intelligence can be so precisely described that a machine can made to simulate it." And so over the course of eight weeks, these scientists gathered to collaborate and brainstorm ways to make machines think, learn, and solve problems. While there was no immediate breakthrough, this conference marked the formal introduction of artificial intelligence as an academic discipline and laid the groundwork for future research and development. Fast-forward to the early 1970s, medical science stood on the cusp of a technological revolution. Scientists sought to apply the field of data analytics to diagnostics. One such example was a computer system named Internist-I.

Developed at the University of Pittsburgh, this groundbreaking project was designed to unravel the mysteries of clinical diagnosis within internal medicine. Aiming to create a system that was equal to or even better than the diagnostic capabilities of seasoned physicians. Another system developed at Stanford University, called Mycin, sought to aid physicians in the diagnosis and treatment of bacterial infections. Both of these early systems relied on the manual input of patient data. While they were able to provide accurate answers, they were never widely used. From these early efforts, AI has evolved dramatically. Today's systems not only analyze vast data sets, but also learn and improve autonomously.

Transforming clinical practice through risk models, enhanced diagnostics and streamlined workflows. In 2021, the University of Michigan became the first university in the nation to create and deploy its own generative AI tools for faculty, staff, and students. Through its e-health, an artificial intelligence initiative or e-HAIL, a collaboration between Michigan Medicine and the College of Engineering, the university is breaking new ground in healthcare innovation. Today's guest is Dr. Cornelius James, a clinical assistant professor in internal medicine, pediatrics and learning health sciences.

He is a member of e-HAIL and the Michigan Integrated Center for Health Analytics and Medical Prediction, or MiCHAMP and is leading multiple projects focused on medical education and AI. Welcome to the show, Dr. James.

Dr. Cornelius James:

Thank you, Kelly. I really appreciate the opportunity to be here to speak about this very important rapidly evolving topic.

Kelly Malcom:

This is kind of a big question, but why is everyone talking about AI seemingly all of a sudden?

Dr. Cornelius James:

I often tell people that there's a history of AI dating back to people like Arthur Samuel, John McCarthy, Turing with the Turing tests, and even more recently with Geoff Hinton. He's considered by many to be the godfather of artificial intelligence. He did a lot with respect to deep learning, but those things were basically, or those people were very productive in say, the 1960s, 70s or so, and Hinton more recently in the 2000s, but I would say it was 2022, November of 2022 specifically, which is when ChatGPT came out or when it was released. That's ChatGPT 3.5. I think that's when a lot of people kind of took note of the potential of AI, and I think a lot of people are talking about it now because AIs more readily or people recognize that they have AI in their hands.

In many cases, they've always had, or not always, but for a long period of time, they've had AI in their hands. You have a smartphone and if you've ever used the Maps' application on your smartphone, or if you've ever looked at the location of where you took a picture or if your email has ever sorted things into spam versus email that you may actually be interested in reading. You've had AI there, but it wasn't, at least in my opinion, until ChatGPT in November 2022, that was when I think a lot of people started to really recognize the potential that AI could have not only in healthcare, but in everyone's lives.

Kelly Malcom:

Is there something unique about ChatGPT that makes it maybe well suited to healthcare compared to maybe some other algorithms or other computer-based programs?

Dr. Cornelius James:

That's a great, great question. As I mentioned, there have been AI systems or machine learning models that have been deployed or implemented even prior to ChatGPT. Many of those were what we describe as what's called narrow AI, and that they performed a specific task. So you couldn't really adapt them to a number of other types of tasks. Whereas with ChatGPT or with large language models, those are being adapted for various things and I'd say more specifically when it comes to administrative burdens that clinicians are facing, when it comes to documentation and reviewing patient records, recording or documenting encounters with patients. It's things like that, but then also using it to assist with clinical reasoning.

So this is where we talk about that more generative or getting closer to what's called more general AI, which is what allows us to adapt or adopt the AI for different reasons as opposed to that more narrow AI, which is often used for a specific task. So when we think about AI in a healthcare setting, right now we're seeing a lot of movement with respect to using large language models, again, for those administrative purposes, which is something that's a big deal because we know that that's causing a lot of stress and burnout for clinicians. Another long ramble-y answer, I'm sorry.

Kelly Malcom:

No, that's perfect. I was just going to say for our benefit, what is a large language model?

Dr. Cornelius James:

So a large language model basically is a type of artificial intelligence. So there's something called natural language processing that's been around for a while, but a large language model is often built upon something called a foundation model, but it's essentially something or type of artificial intelligence that can understand human language and that can also generate human language. So that's kind of just of what a large language model is.

Kelly Malcom:

What are some of the potential applications, and if we're already implementing AI, what are some of the actual applications of AI in healthcare?

Dr. Cornelius James:

I mentioned there's the narrow AI and a lot of the work that I had been doing prior to ChatGPT, it is more so related to those narrower applications of AI. For example, there was a model released not too long ago that got a lot of attention because it was the first AI based system that could autonomously diagnose diabetic retinopathy. So that's an example of, again, that narrow AI where we're using it for a specific reason, and that's great. So we've seen a lot when it comes to AI in the diagnostic space, more specifically in imaging based specialties. So that's again, the narrow AI. Where we are now is again, the generative AI, things like your GPTs and so on, which will again take care of things like potentially... It has the potential to assist or augment clinicians from an administrative perspective.

So again, Ambient.ai, right? So there's AI listening in the room, hopefully with patient's permission, and then it'll generate a document or generate a clinical note based upon what it's hearing in the room. So that's something that has been implemented in some institutions. And then there's also patient portal. So the patient portal is something you have in-basket and clinicians are pretty overwhelmed when it comes to receiving messages from their patients. We all want to take the best care of our patients. We all want to respond to them in a way that's convenient, appropriate, that's going to be helpful for them. But doing that or managing that in-basket or responding to those messages, it can become fairly challenging. And there's something that's called pajama time.

And a lot of time physicians spend a lot of pajama time working on things in the electronic health record, including portal messages. So with AI, there are some systems that have already started to use AI to respond or to generate responses to patient portal messages. Now, clinicians will still have to exercise or have some vigilance and review what the AI has generated, but it does still potentially shorten the amount of time that a clinician has to spend in the electronic health record responding to a patient portal message itself. I think I just gave you kind of a clinical example, but then also an administrative example as well. And those are just clinical and administrative. We also know that there's a lot of work that's being done in the medical education and the operation space.

Kelly Malcom:

So what are some of your concerns or the limitations of AI in healthcare and how do you think we can overcome these?

Dr. Cornelius James:

These technologies are going to continue to improve. They have shown that they're going to do that. When you look at ChatGPT 3.5 versus ChatGPT 4 and ChatGPT 4.0, they just continue to improve. They're going to become more accurate as time passes. The question is how do we implement these technologies or these tools into clinical practice? And not only how do we implement them, but how do clinicians adapt to them and begin to work with them and to collaborate with them or team with them? And that's something that we're going to have to think about from a continuing medical education perspective because there's currently clinicians out there that are practicing, but then also as we bring new physicians and healthcare professionals into the fold or as they begin to care for patients or learn to care for patients, that's another group of individuals that we have to train or teach about how to use these technologies.

So the implementation, I don't suspect that it's going to be the technologies themselves that are ultimately going to be the limiting factor when it comes to how much benefit we derive from these technologies. I think it's going to be more so the workflows, the culture, the systems that we are attempting to implement them into. And we don't want to be sloppy in doing that, but we want to be very thoughtful and careful as we do that. And then there's also the governance piece. I think governance and regulation, we're still evolving when it comes to that. When it comes to... Nicholson Price wrote about something called collaborative governance.

So governance from the federal government or maybe even internationally, and then governance at a state level, governance at the local level, and then governance at an individual clinician level. So all of those individuals ensuring that these technologies are being used in a safe, ethical, responsible manner is another concern or something that I think we have to tread gently in or tread slowly in or carefully in as they're increasingly implemented into the clinical spaces or just in the health system in general.

Kelly Malcom:

We've kind of touched on how AI could improve sort of clinician workload. How can it improve a patient's experience?

Dr. Cornelius James:

Wow. That's another great question. Myself, Karen Deepseeing, Tom Valley and Jenna Wiens wrote a paper, it was in AHRQ, and we described this patient-clinician-AI triad. And that's going to be the reality of patient, clinician, and AI encounters. In that clinicians will be using AI to care for patients. Patients will be, and currently are using AI in some respects to care for themselves, and patients will bring their own AI into the relationship. And again, clinicians will have the AI that they're using. I see the potential for patients to become a bit more independent, which I think is a good thing, but they're going to have to be, in my opinion, taught or trained on how to use or engage with these technologies. We've often heard of the Google patient or people who Google things, and then it's kind of like that.

I've never really had a problem with people that Google things and bring it to me. I think it's a good thing. It's information. The concern is that there's a lot of bad information out there, so there's a lot of bad information. And quite honestly, I anticipate that there's going be a lot of bad AI out there. And I think as clinicians, we're going to have some responsibility when it comes to teaching and/or helping our patients or walking with them as they figure out what AI works best for them. Because you could imagine an algorithm out there that instruct if they... A patient has a certain condition and an algorithm is telling them to do something that could potentially worsen that condition.

And those models are out there. When you think about the Apple App store or the Google Play Store, over 300,000 health related apps. Some of them are FDA approved, some of them are not, which means some of them are regulated, some of them are not. So you could imagine a patient with a certain condition, let's say for example, they have heart failure and there's a certain amount of salt that they should or should not be consuming. You could imagine, depending upon the application that they're using, that app could be encouraging more or less salt than they should be getting. So those are the types of things that we're going to have to kind of think about as patients begin to use these technologies.

I often also liken it to over the counter medications. We're going to have over the counter apps or over the counter AI that we're going to have to address or deal with. And I think very similar to the way that clinicians are trained on the risks and benefits. We can't know every single medication out there. We likely won't know every single AI that's out there, but to try to provide some sort of overarching and/or broad kind of guardrails that'll allow patients to at least take a step back and say, do I want to begin engaging with this app prior to running it by my clinician or my clinician has given me some guidance or some direction when it comes to the types of algorithms or AI I should or should not or may want to consider not engaging with.

Kelly Malcom:

For individuals who may be skeptical or nervous about implementing AI into healthcare and doctor's offices, what are the benefits for the patients?

Dr. Cornelius James:

I often say when it comes to... So I direct the evidence-based medicine curriculum at Michigan, so that's appraisal of the literature. You want to have a good eye toward being able to say, I believe or I don't believe this literature. And you want to be as objective a way as you possibly can. So you have to be trained to be able to determine what has solid internal validity, what's believable. So for that reason, I say skepticism when it comes to AI, I actually think it's a good thing. The cynicism and apathy is what I really like to try to avoid. And I think that's for clinicians, patients, and just people in the world in general. Are there potential bad players everywhere? Absolutely.

And I think we are all going to come into this new technology based AI based healthcare with our own backgrounds and experiences. And as clinicians, I think we have to understand and respect that. I currently use a scribe, but it's a person who is listening in on my phone and I have to ask the patient if they are comfortable with that. And if they're not, then we move forward recognizing that that's something that they're not comfortable with. And I think the same will be true for AI. There will be some that are more or less comfortable with it, but I certainly believe that it's essential that we allow patients to maintain that autonomy and understand, and this is going to be something that we're going to have to kind of think about.

Because some things in medicine, they just kind of happen where there is not that informed or assumed consent or we have to kind of think through that. And I think AI, we're still figuring out when that's going to be. When do we inform a patient that AI is involved in their care or that it's involved in this encounter? Those are still questions that we have to ask, but I think ultimately we have to default toward ensuring that patients and clinicians are able to maintain their autonomy.

Kelly Malcom:

Is there any research being done or evidence that shows that AI tools benefit rural or low resource hospitals and healthcare systems?

Dr. Cornelius James:

Earlier I mentioned that model that could autonomously diagnose diabetic retinopathy. So you could imagine that if that technology can autonomously diagnose retinopathy, if it's embedded into a primary care physician's office and it can say you do not have diabetic retinopathy, then we potentially avoid a referral to an ophthalmologist. So you can imagine that there are some settings that don't have all that many ophthalmologists in that zip code or in that area or in that county. So I don't want to oversimplify things, but certainly models like that can benefit communities like that. And there's people like Akbar Waljee who is a leader and one of the directors of e-HAIL who does a lot of wonderful work in Kenya.

And I know he's looking at AI and the benefits that we can... And he's a gastroenterologist, but he's looking to see how those types of technologies can benefit people in lower resource countries. So one of my mentors often says we have a lot to learn from what's happening in lower resource countries when it comes to adopting and adapting these technologies. And I think that the work that Akbar is doing could potentially inform how we do address and/or get these technologies implemented knowing that there are some differences, but how we get these technologies integrated into lower resource settings. So he's doing a lot of great work there.

Kelly Malcom:

Can you tell us a little bit about e-HAIL?

Dr. Cornelius James:

E-HAIL is wonderful because... And I'll go back to even before describing e-HAIL. I have been privileged in that I've been able to work with a multidisciplinary group of individuals to come together and develop curricula to train clinicians to use these technologies. And a lot of those individuals are part of e-HAIL. So e-HAIL brings together people from all over, right? Law, medicine, research, clinicians, et cetera, computer scientists and so on. It brings all of these individuals together because we can appreciate the broad and far-reaching impact that these technologies are going to have in healthcare and beyond. So e-HAIL is designed to bring individuals together from these different disciplines and to determine or find or identify areas where there's a mutual interest.

Where we can say, "As a computer scientist, engineer, as a developer, I'm interested in this particular form or type of healthcare or this problem." But then you have the clinician and/or clinician researcher that may be able to say, "Well, you know what? That sounds really cool. It sounds really nice, but this problem is even cooler. So can we come together and can we address this problem because this is really what clinicians are really struggling with?" And the computer scientist and/or engineer can come in and say, "You know what? If you want to develop that algorithm, this is what we're going to need." And the physician and/or the healthcare provider can say, "Well, if we develop that algorithm, if we want to implement it, this is what we're going to have to think about."

So you could imagine having those perspectives coming together, and this is what it's going to take, whether it's education, research, operations, etc, is going to take this type of collaboration, multidisciplinary effort to ensure that we're successful in implementing these technologies across healthcare generally. And that's what e-HAIL or programs or initiatives like e-HAIL are seeking to do. The same is true for things like MiCHAMP and Precision Health. It's all about bringing people together.

Kelly Malcom:

So how are we here at U of M Medical School preparing the next generation of clinicians to deal with AI?

Dr. Cornelius James:

So we as other institutions or medical schools around the country, we got a lot of work to do, to be honest with you. I know we are thinking about it. I have the privilege of engaging with a number of leaders or people who are leaders in this space. I have the privilege of engaging with them, but we've got a lot of work to do. I know people like Jenny Sheffield is leading a work group at the medical school specifically to begin to think about how we use these technologies both to improve or augment what we're doing as educators or as a medical school, but then also thinking about how we begin to train clinicians to use these tools. So we're going to have to think about this across the continuum of medical education.

And it's going to look fairly different from undergraduate medical education to graduate to continuing medical education and even thinking about undergrad. So our undergrads, what type of baseline knowledge are they coming into medical school with, and where do we need to improve that knowledge or where do we need to solidify or build upon that knowledge? So when it comes to things that we're actively doing, there are some things that we are doing from a medical education perspective beginning to include some sessions and/or talks or lectures in both the medical school and at least I know specifically about the internal medicine curriculum because we actually implemented or integrated some AI content into that residency program.

But there are things like that that are going on, but there's still a lot of work to do though.

Kelly Malcom:

So what are you currently working on and what's the future direction of your research?

Dr. Cornelius James:

I'm very interested in, or one of the concerns that I've had is around implementation of these technologies. So currently I'm working with mentors, [inaudible 00:25:28], Mike Dorsch, Sarah Krein, John Pate. We are looking at a mobile health application that's AI based. It's designed to... Basically, I gave that example about salt because this is what we're doing, but it's designed to encourage people with hypertension to consume less salt, right? So it's an AI based, what we call reinforcement learning algorithm that's embedded into a mobile health application. So my question is, well, this could prove efficacious or effective, or it may lower people's salt or it may decrease their blood pressure, but what's going to get them to use it?

What's going to get clinicians to prescribe it? What are the facilitators and barriers to patients saying, "This is actually an app that I'm interested in using that I'm going to start using and that I'm going to continue using." And then the clinician on their end, is this an app that you would recommend to your patients? What are the facilitators and barriers to you recommending and continuing to recommend this type of application to your patients? And then once we figure those things out, then how do we put some things in place to facilitate the actual use by patients and the recommendation by clinicians and actually getting patients set up with these types of technologies.

And then not only that, but doing so in an equitable fashion. So those are pretty much where my interests lie is digital health technology, including artificial intelligence and machine learning based technologies, and what are the facilitators and barriers to effectively implementing those things, ensuring that clinicians are using them safely and responsibly. The same is true for patients, but then also making sure that they are equitably adopted by people from all walks of life.

Kelly Malcom:

The Fundamentals is produced by the Michigan Medicine Department of Communication in partnership with the University of Michigan Medical School. Find us and subscribe wherever you listen to podcasts.