AI in healthcare and the future of medicine

Episode Transcript

Behind every case, there’s a story. Behind every story, there’s a lesson. And behind every lesson learned, there’s a trusted mentor. This is Prepped and Draped with Dr. John Kaufman, where candid conversations, actual cases, and bold new professionals shape medicine.

Welcome to Prepped and Draped. I’m John Kaufman, the host of the series, and I’m really delighted today to have as my guest, Dr. Judy Gichoya, who I think you’re going to find an absolutely fascinating individual who is an active interventionalist and also a leader in informatics and AI. Judy, can you just introduce yourself?

My name is Judy. I’m an interventional radiologist from Emory University. When I’m not in the IR angio suite, I’m a data scientist and an AI researcher. I run the HITI Lab, which is focused on building AI datasets and working on bias and fairness in algorithms, looking how AI works in the real world and building capacity for the next-generation data scientists.

One of the things that when we first met was really fascinating to me was your journey from just being, what I like to say, a regular person to someone who’s highly, highly respected in the AI informatics world. And I don’t know if you could just tell the audience a little bit about how you did that path, because so many people are becoming aware of this and wondering, “Is that something I could even consider doing?”

Yeah, no, thanks for that question. And so my journey was a little different. I’m a foreign medical doctor, and I went to medical school in Kenya in a university called Moi University, which is actually known if you know some of the— I think most recently, the New York Marathon, the Kenyan runners live in this region of the world.

Cool.

Yeah. And so during medical school, something amazing was happening. And it’s interesting because I think also right now we are undergoing an amazing transformation, especially I believe with what we’ll be discussing around these things called foundation models. And that time during the HIV pandemic, the doctor there, Joe Mamlin, was trying to figure out, how do we organize the medical records to take care of patients in a pandemic? Now, in the past, I would have to explain the significance of that, but hopefully most of us will be listening to this have and remember COVID. And so you have lots of data, lots of patients, and you need to figure out how to prioritize. And so Joe’s son, who is a physician, and actually Joe and Bach live in Indianapolis, so quite some connection I think with some of the stories around this.

Oh, wow.

Are both physicians, but Bach is a computer genius and he was visiting his dad and his dad said, “You have to help me.”

And so when I was a medical student, there was this open-source medical record system that began in Moi University and the AMPATH care system. And so I got to learn how technology can be used to really organize medical records. And so that entry just really sparked an interest and I kept the interest going till now. I have had this lucky touchpoint when Geoffrey Hinton said, “Let’s stop training radiologists.” Now, the foundation models, I think have just been really always at the right place, right time, but also with a very, very strong support network that has allowed me to explore these topics.

And the connection to Indiana, just so people know, is that you did your residency training there and then that’s before you came to us out in Oregon. It sounds like your journey into this was very much related to a real-world clinical, tangible need. You had a tremendous amount of information. Talk a little bit about the significance of open source. I think this means that you were able to access a medical record and start using it without having to go through huge licensing issues dealing with big companies. Maybe just talk about how you feel about open source.

Actually, I’d like to comment about two things. One is this real world and Joe Mamlin, who was the physician then who started this fantastic program called AMPATH, caring initially for HIV patients, always said you have to lead with care, lead with patient care. And this I think has been a very, very good anchor, even a true north. And also why most physicians actually stop practicing, and probably one day I will, but I find myself every time I’m anchored in clinical care that it just reminds you of the problem and what you really need to solve because it’s easy to slap technology to a problem that doesn’t exist, so that’s one end.

Now, open source is not just the licensing, which is important, but it’s the ability to create a community, and anyone who knows me or works with me, I can see that those principles and just having this sense of belonging I think has informed even how I do my own research. And so working with others is amazing and that’s open source. So you create this environment for working with others and it’s not just the technology, but also the infrastructure that is necessary to build a community. And also, this just sense of belonging, your gigs, you end up enjoying just solving problems. We see now even with this competition of large language models that open source has really been phenomenon in trying to even give people a sense of what this technology can do and you’re not just limited to the big tech company. I’ve made a lot of friends over the years and just been able to travel to new places just because of this community that is united with a shared purpose.

Something you just said, because I think it’s so important in this world that we’re living in now that we have our phones and you sit in a waiting room of an airport, which I tend to do a lot, and everyone is looking down into their hand. No one is talking to anyone else. No one is looking at anyone else. The digital world has the capability of actually being very isolating. But what you’re talking about, it is actually bringing together people to have more human-to-human interactions. And I’ve heard AI described recently as the role is to allow us to spend more time being human to human, direct interactions and less time having to do drudgery or just tasks that take a lot of time but are very isolating.

And you just described that perfectly. The journey for you, it sounds like has really been about people and how this has enhanced your ability to interact with people rather than just being lost in the matrix, so to speak, the digital world. You mentioned foundation models. Could you talk a little bit about that?

I heard recently a casual comment that, “Oh, 2026 is going to be the year of the foundation models.”

And I said, “Huh, what?” So maybe you can explain that.

Yeah. So if you think of a building or a recipe and you start off with a structure, so you can have a foundation, the building foundation. And if you are able to prototype, you could maybe say, “It doesn’t matter whether I build a church, a school, a hospital, an airport, that maybe I could have just this foundation that is prefabricated and everything I put on top of it just appears to be different.”

The same concept applies to this change in paradigm, which is saying, “Look, in the past, what we did was build AI to say, let’s predict, is there bleeding or not? Is there a stroke or no stroke?”

So you pulled some CT scans, you ended up saying, “Here are some positive examples, here are negative examples.” You gave it to the model, and the models, especially in the era of the deep-learning networks, were able to learn the patterns that were characteristic of a patient with stroke or no stroke.

Now, the foundation model is the paradigm of you say, “Look, I’m going to get all the cities in the world.” So the most familiar to most of us will be the ChatGPT, large language model partnership. And so you get all the text in the world, and if you are able to load it on compute and process them into these words, then you can start to have a foundation that understands language because you could see maybe when I pull everything that describes Judy and what she posts and what she publishes, that there’s always interventional radiology. It starts to get a sense that, yeah, Judy and interventional radiology may be closer together more than Judy and a pilot or something else. And so you get this foundation, then when you have the foundation, you can build things on top of it.

That’s why you’re hearing this is the year of the foundation because when you have a computer system that has seen all the world’s text, at least in the public domain, then you could start to say, “Write an essay,” or you could have interact with the foundation with a question-and-answer answering system.

Or for radiology, you could start to say, “If you’ve seen all the images in the world, describe these findings,” or the radiologist could say, “Is there a pneumothorax?” So you have this broad-based fabrication infrastructure by relying on the foundation models, but they rely on having a lot of datasets and a lot of compute, which is why they can only be built in very specific areas and all of us tend to be downstream users for them.

I think you just answered my next question. This is not going to be a science fiction, there’s one foundation model for all of knowledge and all of human use, but they’re going to be a little more specific for certain areas, is that what you’re saying?

Yeah, I think because there’s an answer is to build these models because people are saying now we’re moving to general intelligence that the— ideally is that if, John, you could replicate Johns all over the world and they understood all these things about interventional radiology, is that unlike you who gets tired, they could keep working and working and come up with new catheters and new techniques. So people are anticipating that the future is that they could have their own intelligence and work that out themselves. But the paradigm that we are seeing now is really that we no longer just provide the datasets and the task.

Let me step back. In the deep learning era, you said, “Here’s the dataset and here’s the task that I want you to do.”

In the foundation model era, you said, “Here’s the data, get a better sense of the world,” and I can come back and decide what I want to do on top of it.

And now we have the agentic world that is saying, “I’m just going to say, make a podcast or write something. I’m not going to tell you how or what the data is.” And that’s the paradigm that we are in. And so when you think about those three senses is that people have hoped that they will have the one model that does that because it gives you competitive advantage. You lock people into one ecosystem. But the reality when we use this with these models is that there’s no one size that fits all and you still need multiple.

No, that’s a really important concept of there’s a biological variability and there’s always change, and so there’s an infinite variety, and it seems unlikely that there’ll be just one model that will work for everything. And it leads me to my next question to you, which I think many people struggle with, is you live in two worlds in a sense, you live in a procedural world where you’re dealing with patients who have nerves and things wrong with them and you have to do things with them, and it’s very tangible. It’s not a digital twin you’re working on. It is the real thing. And then you have this other world, which is very much a digital kind of world. How do you balance that? And you’ve alluded to this, but talk a little more about how you use those two things to enhance each other.

I’ll start with a little bit of an oxymoron, which is that I don’t believe technology actually impacts patients’ lives. That’s a good reality of your segue to this question, which is that tell me one system or application that if I hadn’t used it, this patient would not be alive. Now you could say, maybe I reduced my medical errors in the era of EMR, electronic medical records, but we know that they’ve also had quite a lot of unintended consequences. Now, that being said, the opportunity to think about technology can be very transformative for patient care and patient outcomes, and that’s a little bit of my enjoyment of living in these two worlds.

For example, at Emory, we have one of the biggest outputs for our team is the Emory breast radiology dataset. And it may surprise people, because it did to me, that for all the movements, the political movement that is breast cancer, the cancers that are caught based on screening, not because of anything else, contribute to 14% of all cancers. That’s a bad test, in my opinion. And so if you come back and you start to look at the infrastructure that is necessary to think about technology, the dataset curation that is necessary, you can ask this question again. And that’s been one of my change in process, which is that for many, many bad things that are possible, AI is a very good feature detector.

You can put datasets and you can clearly see, “Hey, I see this one cohort or one subset of patients are not doing well.” And if you use that technology to study what I would say an old problem, like an unsolved problem, I think it can be very transformative. And so what I believe in terms of the opportunity— and we’ve seen it at Emory, because now we’ve spent all this time curating this dataset, we are supporting our cancer registry. The cancer registries tend to lag two years behind, and that’s the ones that are used for policy.

This concept of data science and organizing these datasets can have a profound impact on patients and their outcomes. I try to remind myself, and as you know, I hang around Dr. Newsome to remind myself why these things matter. For those who are not familiar, Dr. Newsome is another interventional radiologist. And for a long time she’d say, “Okay, why should we care about this?” And it’s always being centered on how does this make our lives also not just patients, but also the physicians and the healthcare workforce that takes care of patients, but also to really move their needle. And that noise can really remove that type of focus.

I’m going to change gears a little bit. Let’s talk about some challenges you’ve had in your career. And I read an interview recently, just to frame this with you, which was a really impressive interview in which you talked about an article that you had written in 2022 about the ability for AI from imaging to determine things like ethnicity of the patient and how that intersects with your interests in applications of AI. But it seemed like that presented some challenges for you, that article.

I like to call it the Reading Race Project. And if I step back out and I had a time capsule, and when I was at the Dotter Institute and put a note with you and said, “Okay, let’s open this on December 2025,” I would not have envisioned that this is where I would be in terms of what I’m doing, but I’m a scientist, I love to do research and ask questions. And as you know that I always anchor with what was happening. 2020 was a lot of problems, not just from COVID, but also the reckoning of healthcare disparities.

And I know that 2025, we are renaming those things to be something different, but a lot of people were saying, “Hey, am I a bystander, or what am I doing to make sure that everyone can get access to healthcare and better outcomes?”

And so this project started off as an ignorant academician. I said, “Yeah, if we have more diverse datasets, we are going to solve all the problems of AI and bias,” and it turns out that we can’t. And so this was a happy accident.

And during our preliminary meetings, actually Po-Chih, who’s a good friend of mine who lives in Taiwan said, “It’s because these models can encode race.”

And I said, “No, you made a mistake.”

And this project taught me a lot of things. One, again, I relied on my open source exposure to build a team that can answer very, very difficult question. When you say that you can tell from a chest x-ray whether the patient is Black or Asian, then you have a problem because you can actually have to justify that this is not a biologic difference. This is just how people see you or how the legal system sees you or how you see yourself. And so to be able to navigate this type of research, I can tell you that I did it in a very naive, but just from a curiosity point of view. But now that as an immigrant I understand quite a lot of history, I think that it would’ve been probably almost impossible to ask that question and study it, and knowing some of the assumptions and the disparities and some of the challenges that especially African Americans face in the US.

Well, we are constantly learning about ourselves, right? We think we understand ourselves pretty well, and this may be an instance of which your AI has unmasked some of the things that we don’t even think about. And I wonder when I read that, what AI or machine learning, all the things, I can’t say to them the way you say it, and in many ways reflect how our brains work. And that’s what becomes shocking to us when we see that this inanimate thing can actually produce material that seems like it came from someone’s mind because maybe we may be wired in a more digital way than we realize. There’s some parallels in how data is managed by AI and how we manage it. That was just really very interesting to me that the kind of conundrum that that conclusion brings up, and how do you then explain that?

Yeah, just that last sentence, how you explained that, because nobody teaches you how to communicate your science. There’s so much of the hidden curriculum that I’ve learned over these years, even as faculty. For example, to be able to publish that paper, we had to release the code and make it very easy for people to repeat experiments. This was discussed on Reddit, there’s quite a lot of debate about it. There was also some great comments. Some others were very toxic. This is the internet era, and it just really taught me how to think a lot about how you communicate your science. And along communicating science is also the humility to say you don’t know or you’re working on it.

And so I found that luckily this still has stood the test of time. We still cannot say what the one thing is, but it’s an area of science and research, but that was supported by reproducibility. And that is extremely important today where we are having quite a lot of mistrust with science. It’s not just enough to be in our academic halls and publishing these high-impact journals. It’s also how we move and communicate and engage the public with science and that’s a big component of my work and consideration for outreach.

Well, this has been a great conversation. And as always, you may not think of this, but I’m the student and you’re the teacher, and I love that.

Good thing it’s being recorded.

Exactly. Just in closing, any last thing you want to say? I know you’ve mentioned many times students and working with students and things that you’ve learned. Any last thing you want to mention before we close out?

One of the things is that for a long time— and this is truly, I’m a scientist, but I love students and that’s my superpower. But I know that you do because you taught me, and I got this experience. And you may not remember this, but during around three months, I was very frustrated about the palpation access of arterial system, the femoral artery by palpation. And I kept asking and I felt very safe to ask what would be considered stupid question like, “Oh, but I’m hitting the bone.” And I remember that I was very, very frustrated.

And you said, “Oh yeah, that’s what’s supposed to happen.” And it reminds me as someone who trained at the end of an era for how IR was taught and how IR really was performed, that as educators, we really need to think about the mental models. I think our specialty, both these worlds that I live in love to complicate things. And if you could spend the time to figure out how students learn or help them figure out how they learn and their mental models, it can be very transformative.

It is really an amazing, amazing, amazing time, not just because of AI, but just because of just society and some of the greats of medicine. For example, when I was a student, I didn’t understand why— the role of immunology, but immunotherapy is so transformative. And our IRness is that we live outside just our suites and you could be touching a toe in the morning and the chest ports in the afternoon and the veins in the evening. And that IRness that I can never really describe, it really positions us to live in more than one world and can be very transformative with patients. But if we keep making it very difficult for people to participate and forget those fundamental foundations of IR, I think that we have a missed opportunity, and so I hope that some people will get inspired to do that.

Oh, Judy, thank you very much. There’s no way I can say anything to add to that, so really appreciate your coming on and being a guest today, and I hope the listeners found this as fascinating as I have.

Thank you, John.