The Future is Now: Ryan Stidham, MD, on Practical Applications for AI and ChatGPT

In the latest IBD Drive Time, host Raymond Cross, MD, talks with Ryan Stidham, MD, about potential applications for artificial intelligence and ChatGPT in gastroenterology practice and managing inflammatory bowel disease—and what is feasible right now.

Raymond Cross, MD, is a professor of medicine and director of the IBD Program at the University of Maryland School of Medicine in Baltimore. Ryan Stidham, MD, is associate professor of gastroenterology and hepatology at University of Michigan Medicine in Ann Arbor.

TRANSCRIPT:

Raymond Cross:

Hi everyone. Welcome to IBD Drive Time. I'm Raymond Cross from the University of Maryland School of Medicine, and I'm delighted to have my friend Ryan Stidham from University of Michigan, as our guest today to talk about artificial intelligence in GI. Ryan, welcome to IBD Drive Time

Ryan Stidham:

Ray. Always a pleasure. Thanks again.

Raymond Cross:

So for our novice listeners, which includes me, is there a definition for artificial intelligence? What does it encompass?

Ryan Stidham:

It's a term that's stuck, but it's not really a good descriptor of what it is, but I'm not going to go on a nomenclature war, but we'll just stick with it. So what's meant by artificial intelligence is really two different things. It's a set of tools that are able to extract information from a variety of sources and then apply a variety of different analytic methods to attempt to infer some kind of judgment from that. Most of those judgments that are inferred are often directed by humans or by some kind of known outcome. So in our space it may be is a medication going to be successful? Is it going to resolve symptoms? Will the patient end up having surgery? Will they have an allergy to the medication? as a few examples. And so this series of technologies that go into that is actually a bucket of a bunch of different things.

So you begin with access to the information itself. And in the past we were limited to so-called structured information such as spreadsheets and tables and lab data, administrative claims data, and we had the analytics for that. We had a variety of logistic regressions, which are statistical models, but then increasingly so-called machine learning models, where the architecture and weighting of the different variables in the model can change and the model can adapt itself until it optimizes and gets the best possible result. And so that was traditionally where we were at for a long time, but what's changed recently is the computational power has improved so that a lot of these more sophisticated models can now start extracting information from so-called unstructured data, things like imaging, where there's an enormous amount of information present, but often we're trying to lump it into one group or another. Is there disease present or is it absent? Is it severe or is it mild?

And the technology now exists. In fact, the methods have been there for close to 40 years, but the computational power wasn't really there to enable those methods. And so we're now extracting information from these unstructured sources like imaging and text and starting to use that in our models now, which is really part of what's powering this AI revolution. Whether it's the automatic vacuum in your home that can see and adapt to the environment or in our case if it's models trying to tell us if we should escalate a drug or give up and move on to something else. So in summary, that's what we're dealing with.

Raymond Cross:

You talked a lot of predictive tools, which we clearly need precision medicine, right? We've been talking about it for a long time. We don't have it, we're not really even close. It's all clinical factors we still use, but I can think of other areas where this could be incredibly helpful. So for example, when you get an endoscopy report from an outside physician, how they describe what they see and whether they put a quantitative score on it, is vastly different. And likewise for cross-sectional imaging, one radiologist reads differently from another radiologist. So you can see that there is a great need there for some standardization and something that could automatically interpret those images and say, this is what's there. So those are a couple examples I'm thinking of, but are there other areas, you talked about predictive modeling, are there other areas I'm missing and are we making any progress in IBD?

Ryan Stidham:

There's a lot of progress that's being made. I mean most of what is near term in terms of what's going to be available, you've kind of hit on already, which is a lot of the imaging-related work. Because the various neural networks that are now available and other things called adversarial networks, those details are less important unless you want to know about them, but they're extremely good at image analysis because imaging is really just chunks of categorical or even binary data that, if you had enough time, you and I, Ray, could break down a CT scan analytically. It would just take the rest of our lives to get through one 700-slice CTE, and the machine will do it very, very fast—before I can finish this sentence.

So in the imaging space, everything you said is totally true. You get more reproducible results, you get more granularity, you get more quantification and extracting the factors that we know as well as perhaps unknown disease factors that seem to be correlated with one state of disease versus another, medical response, nonresponse degrees of severity, degrees of symptoms, et cetera.

But there's a lot more than imaging. I think a lot of the focus is on imaging is because it's something that the field is really good at now and frankly there's a lot of commercial potential for it. And so at the risk of sounding pessimistic, money drives a lot of things and there's a clear market need for that. So that's where a lot of effort has been. But the other spaces are text. And we can talk about applications for this, but these concepts of taking information and training machines to understand it the way that we do, has been very difficult to do in the text space, so-called natural language processing. But different methods than neural networks have now been developed, that are allowing software to really understand the context and tone and meaning of different statements that you'll find within text, with the ability to adjust for the contextual situation.

If there's say an outside referring physician document that's talking about erythema nodosum, well sure you don't need advanced technology to tell you if erythema nodosum is mentioned in a document, but you need a lot of technology and inference to be able to understand, well, is that referencing this patient's symptoms, a family history of that particular issue? Was it in the past? The present? Am I worried about this in the future? So all these subtle contextual things that human comprehension can determine quickly, we're finally getting to the point where machines are able to do this. And so this whole longitudinal record that we keep, this diary of patient experience and symptoms and the journey of a patient through IBD that we've logged in all these electronic notes now, is finally extractable and can be organized. So in that sense, all of the nuanced detail that is in your notes and in all of our documentation that just isn't captured by the labs or the diagnostic codes that we give the patient or the outcome doesn't quite capture everything, these are now pieces of information that are possible to use.

So it's not that the models are getting better or that there's some better supercomputer that's more powerful or smarter. It's instead, the models themselves are actually the same. The model to predict disease progression, the model to predict adverse events, the model to predict likelihood of surgery or therapeutic success, those models are actually still quite the same. What's new is the automated capability to organize and structure all the information that previously was just not able to be collected at scale. And so the imaging that you talked about plus the natural language processing, are 2 of the big things that are going to have many, many applications in the near term for both predicting course, getting us closer to precision medicine. Because If you have better data, you're going to have better models. And then also some just useful tools that'll make our lives better.

Raymond Cross:

Well, hopefully when the computer's analyzing what I put in my notes, they're not going to spit it out and say that Ray Cross is stupid. So that would be not optimal. You're talking about imaging. So I wanted to talk a little bit about intestinal ultrasound, which is an extremely hot topic and you and I have had a few discussions about this and I think a couple of challenges relate to is it giving us anything more than what we have as far as improving patient care? But some of the other issues relate to the training that's required, the cost of the machine and the cost of the probe. And I was thinking, you would think that artificial intelligence or whatever we're going to call this, would be perfect for something like that. Why couldn't you have something on your phone to acquire the images that could then be interpreted in real time? And do you think that's where that's going? If it's proven to be additive, do you think that that AI could help with that?

Ryan Stidham:

Yes. In fact, if intestinal ultrasound proves to be useful at all, we will need artificial intelligence to actually make this practical and useful for everyone. We'll put to the side the question of the utility of ultrasound itself. I'm still skeptical. I want to be proven wrong. I would like to see this work. But the point about can AI help solve these challenges, intestinal ultrasound is the perfect, perfect problem for AI. Because it fits in this 3-part Venn diagram of where is AI useful? Is the task possible by humans but requires expertise and that it's difficult to learn? Is a task something that we don't want to do? And is it achievable? And so I think that this intestinal ultrasound problem fits very perfectly in there. You can imagine a future where, what's one of the barriers of intestinal ultrasound? Well, you have to be trained to do it. And even when you're trained to do it's still extremely difficult.

Raymond Cross:

It's a point of care test too. So you have a 20- or 30-minute appointment and how are you going to integrate that into the appointment? If it's not done in real time when you're in clinic, then to me that's less valuable. So we need something to make it faster to acquire and to analyze.

Ryan Stidham:

So the direction that you go with that as an example problem, is you say, wouldn't this future be better? I see the patient point of care. I more or less wave a wand over that patient. I don't even know what I'm looking at. I just need to have myself or someone else just acquire all the images. You take your linear probe and you are just instructed to do a number of overlapping passes over the anterior and lateral aspects of the abdomen. And then you change modes to, if you're using a radial probe to a deeper setting, do the same thing. You're not even looking at the screen. And then you allow the machine to take those images, more or less identify what are useful images, what are not useful images. So you do a informativeness classification. You then have to do an image segmentation where all those images, the machine has to be trained, what's bowel, what's not bowel, what's distractor, what am I not sure about?

And then you have to manage the dynamic aspects of the examination. The bowel is moving, it may not always be in the same place that you expect it to be. Diseased bowel is easy to find, normal bowel is very hard to find. And sometimes if you've got someone who's got a stricture and they've got a lot of dilation, you have to disambiguate, is that dilated bowel or do they have to pee and is their bladder full? Because the two are adjacent to each other and can look the same? These are things that machines are perfect for. Ultrasound should not be done by people, the technical part of it, unless they are very, very highly expert. Because there's too many scenarios that are difficult to account for. So for all of these reasons, AI is going to not only be helpful, but my opinion absolutely necessary to realize intestinal ultrasound as a practical tool in clinic that anybody can do. And I have no doubt that industry is all over this and this would be something where they can really help change the state of care as well as academic groups.

Raymond Cross:

Yeah for sure. Before I ask you a few more questions, I just want to remind the listeners that we are sponsored by Advances in IBD and the Gastroenterology Learning Network. We're also available on Spotify and Apple Podcasts now. So look, search, Gastroenterology Learning Network, and IBD Drive Time and you'll find us. And then lastly, the in-person national Advances in IBD will be in Orlando December 14th through December 16th. There's plenty of time to register for that wonderful course, which is I'm sure Ryan and I's favorite.

So let's talk a little bit about ChatGPT. So I was listening to a podcast about this. It's interesting, also heard a quote that was used that when we found fire, clearly there's good things with fire and bad things with fire. And so with ChatGPT, I think that's a good analogy. But there's a number of issues as academicians. So you're grading coursework, manuscripts you're reviewing that are submitted to journals, grants. You could see that this technology could be abused. So what should we do? What are your thoughts on this? And when you're reviewing things, are there any techniques that can help mitigate this?

It's troubling.

Ryan Stidham:

There's a lot of places to go with this, and it depends if we want to go to a happy place or a dark place or somewhere in between, because you can find all of those With GPT. Where do you want to go? You want to go to all of them? Where do you want to go?

Raymond Cross:

Well, I think that the bright place is if you're sort of stuck and you need an outline, a template for what you want to write, to me that's a really good use of technology. It helps organize your thoughts. And then you can write and add in. Where I was going is, I want what I'm reviewing to be someone's original work and so how do I mitigate against that? So I guess I'm going dark.

Ryan Stidham:

Well, we could go way darker, but on that particular topic, because there's a lot to say about GPT, so in 1970 at the University of Michigan, for introductory mathematics courses, you were not allowed to use a calculator. Calculators became available. Texas Instruments was making them available. You couldn't use them because it would seem to be cheating or an unfair advantage, or they want you to learn these skills. We wouldn't think twice about using a calculator now. Why would you spend your time doing something that's purely mechanical that a machine can do, and you use your brain to come up with the ways to solve the problem? And I think that the positive view of this as GPT in terms of scholarship and academia, is a lot like a very sophisticated knowledge calculator. And the human job for the coming 200 years is really going to transform to being a knowledge economy, where our job is to learn a bunch of knowledge and facts and then recapitulate them. But we're going to live in an era now where the machines will do all that for us. And our job is to be a creative engine.

I don't really have to care about doing the work to look up my references anymore, or to providing the background. I can have that done mechanically. But it's my job to review it for creative content and nuance and what are statements that are really meaningful and valuable and a new insider impact for the field in terms of our scholarship? So it doesn't feel good for all of us who have spent months writing papers and doing all that research and chart review and literature search. It doesn't feel very good to have that work suddenly boiled down to a couple hours of time on GPT.

But I think that having that tool available is really going to have to, it will elevate the quality of what we start seeing published in our journals. Because if anyone is able to just rehash all the facts out there and slap it together in some kind of review article with a couple of superficial insights, that's no longer going to be of interest to anyone. You're really going to have to push human beings to be more creative. And all this knowledge, information acquisition and organization, the machine will do. But I think instead of all of us, Ray, spending hours and hours and hours on our references and background and introduction, et cetera, we're going to spend hours and hours and hours thinking and talking about what are the real problems we have to solve? What are the creative ways to solve them? What ways really haven't been tried? And so I just think it's going to change how we spend our cognitive time.

So the University of Michigan is actually going, as an example just because I know this one, is going to be allowing GPT use in most of its coursework. Because they see the future for our students of being a world where if you're a financial analyst and you're being asked to review 20 different companies and provide reports, of course you're going to be using GPT. And the same is going to be true, not just in academia, Ray, but in our clinical practice. All of us have increasing pressures in particular with IBD, to really manage and care for more and more patients. It seems like the volume is just really increasing.

Raymond Cross:

Gosh, if we could have a GPT scribe …

Ryan Stidham:

Oh that's no problem.

Raymond Cross:

Our life would improve dramatically. So we're not obsolete. We can see more people and just not have to come back and do two hours of charting.

Ryan Stidham:

That's a great segue to a practical, a very practical use of GPT. Now I have nothing… I wish I had disclosures here. I have none. But as some examples of technology that's really going to help people. There's a system called DAX, which is from a company called Nuance. Are you familiar with them?

Raymond Cross:

No.

Ryan Stidham:

So Nuance is the company that's in a lot of smartphones and EMRs that does the voice to text transcription. And DAX is this system. It's a piece of hardware that goes in the clinical office, in your patient exam room, and it listens to the conversations in the office. It knows who the patient is, who the caregiver is. Another person in the room who might be patient adjacent like a family member, or a person who may be clinician adjacent such as a medical assistant or other technician.

The system DAX will record and transcribe the entire conversation between everyone and then a form of medical GPT—and I'll be actually showing this at the Advances in IBD conference it turns out—will take that entire conversation between you and the patient and the family and the medical assistant and the nurse and you can just ask it, "Can you just write me a note summarizing this?" Done. You'll look at that note and say, "You know what? Could you summarize the orders that I wanted to have?" Done. And maybe you didn't talk about an order. Maybe you wanted a CTE or an MRE, but you forgot to mention it. And you'll just add that in there.

And that's not future coming soon. That's more or less, right now. So yeah, so that's an example of how these tools can be helpful for very practical things.

Raymond Cross:

Well, I mean, Ryan, this is really helpful. I am sure the listeners learned a lot. I learned a lot. Before we let you go, and by the way, we'd love to have you back in the future, tell the audience something about yourself that they may not know or something maybe that I don't even know.

Ryan Stidham:

I don't even know. I'm not very interesting— like personal?

Raymond Cross:

It could be whatever you want. We've had people disclose that they're musicians. David Rubin is a fixer-upper at home and redid all their closets during the pandemic. I make pickles.

Ryan Stidham:

So I do a lot of work in image analysis and a lot of things with video and 3D imaging. And the reason that I'm in that space and applying it to IBD, is because for a long time I did animation. So I did computer animation back in the 1990s, so the early 1990s. So I would do 3D animation for commercials and TV shows. And this was at a time when it was more difficult than it is now. A lot of command line interfaces and hard coding Unix servers and things like that. But I was supposed to take my dad's company, which does industrial film and video production, and that's what I was going to do. But I became more interested in medicine and kind of fell into IBD and fell in love with it. And so a lot of this imaging and tech and AI work I've done is really from stemming from all my days in animation and video and imaging as a kid and teenager,

Raymond Cross:

Exactly what we're talking about. I did not know that. I'm glad I learned that. Ryan, this has been fabulous. Thanks for doing this.

Ryan Stidham:

Ray, always great seeing you, man.

Raymond Cross:

Likewise.