This transcript has been edited for clarity.
Eric J. Topol, MD: Hello. This is Eric Topol with Medscape's Medicine and the Machine podcast. We have a new episode today with a really interesting fellow physician, Dr Adam Rodman. He's an internist at Beth Israel Deaconess Medical Center. He's an educator and instructor at Harvard Medical School, author of the book Short Cuts: Medicine, and has a podcast called Bedside Rounds. He also had articles in both the August 3 and August 10 issues of The New England Journal of Medicine.
Adam Rodman, MD, MPH: Thank you. Eric, I read your book years ago. I'm a little starstruck talking to you.
Topol: You've been really lighting up lately. Before we talk about medical education, which is where you are leading the charge, I thought we would start with where we stand with the large language models (LLMs): ChatGPT, GPT-3, GPT-4, Gemini, and many others that are out there already. How do you see these becoming rooted in the daily practice of medicine? We saw early on the hype about the US Medical Licensing Examination (USMLE), but then we got to patient questions, front door for the doctor, clinical reasoning, diagnosis, etc. What is your vision for where this is going over time?
Rodman: In addition to education, I do clinical reasoning research. There are a couple of domains. I'm sure you saw the preprint from Stanford looking at summarization of documents. We generate a lot of text. Much of what we do is summarize text. My job as an internist is obviously collecting information from patients and building relationships, but it's also absorbing a lot of information that's stored in the chart. Much of the near-term uses that we're seeing, and what people are building right now, is text summarization.
You mentioned clinical reasoning. Most older studies on clinical reasoning were done on the MedQA database for benchmarking, the USMLE questions or USMLE-style questions. My research has been a bit more in-depth about what it means for an LLM to show clinical reasoning or to act as clinical decision support (CDS).
We've actually been thinking about this since the 1950s, but we hadn't taken it too seriously until a year ago. Between 1995 and 2019, at least 15 papers were published evaluating CDS for diagnosis. In the past 6 months, there have been four times that number of papers. So, clearly interest has exploded. This is exciting. If you look from a cognitive psychology standpoint on how doctors make clinical reasoning decisions, a lot of it has to do with words. It has to do with script theory, with semantics, and the understanding that our internal map of diseases and disease processes is mapped out in association with these different semantic qualities.
LLMs don't work the same way as the human mind, but the evidence is starting to show that general-purpose models appear to make medical decisions, or at least diagnoses, equivalent to or even better than humans in very specific experimental situations. It's something we haven't dealt with before in the long history of CDS.
Topol: Well, hopefully everybody is up on how CDS has been evolving. Whereas before it was heuristic primitive algorithms, now it's becoming multimodal with the foundation models. It's much more sophisticated than it used to be, when doctors didn't want CDS in the EHR because it was often stupid or unhelpful. Now it has a chance to really make a difference.
It reminds me of a classic paper that Jerome Groopman wrote in The New Yorker about how doctors think. He even wrote a book about it. You wrote in NEJM that technology has disrupted the way physicians think. Can you elaborate on that? You go back to the Hippocrates era.
Rodman: In that piece I drew two examples: the physical exam (or the pathologic anatomy) and the EHR. If you go back to the beginning of what we generally consider modern medicine — Paris in the late 18th or early 19th century — you find the nascent idea that when someone dies, you can do an autopsy and you can figure out that there are different changes in the body. Like Morgagni, the Italian. When his patients died, he would cut them open and say, "Oh, look, there's a bunch of cheesy material in the lungs. That's weird. Maybe that's why he was short of breath."
Then in post–Revolutionary War France, a movement took place among doctors who began to experiment with what we would now call diagnostic tests. The first was percussion, to see where fluid was accumulating — percussion of the body cavities. Then you have René Laënnec, a musician who used a lathe to make a wooden cylinder, inventing the first stethoscope.
The idea was that you can detect disease on the inside of the human body. These changes happened over a generation, but it was absolutely fundamental to our understanding of disease and the role of the doctor. We take our role as disease detectives for granted now. But the concept that we would have to investigate to find the source of disease rather than it being self-evident is predicated on this idea that we can use tests to figure things out. This whole idea of clinical data — my hair is going gray — but if you go back a generation, no one would have used the phrase "clinical data." We would have talked about facts of the disease. The idea that there are data out there to collect and curate and use to find out things about the patient, all of that tracks back to the development of this tool of diagnostics — the physical exam.
The second big change was the EHR. Most people have a sense that the EHR came out of the informatics movement. It's fundamentally tied to our understanding of AI. Larry Weed, who developed the problem-oriented medical record as well as the SOAP note, had this idea that physicians needed to collect and curate information so that computers could read it and make better decisions. That influenced our work. A few generations of doctors ago, your progress notes would be super-short. They'd say things like "The patient was sleeping; they're feeling better. They had a fever. I had to start ampicillin."
Progress notes became very long, unfortunately, for Larry Weed. He had a lot of regret toward the end of his life. The idea of doctors being data curators came to pass, but CDS and AI, which would take better care of patients, did not. So, we've reframed the way we think about our jobs in view of these technologies.
Topol: That makes you unique because you're a historian of medicine and, at the same time, a futurist. You've made a good case for how this technology can revolutionize or transform clinical reasoning. It's like another chapter. The other thing you touched on, which I thought was really interesting, was Laënnec's invention of the stethoscope, which obviously became a pivotal tool for physicians. As a history-of-medicine expert, you are aware that the war against the first stethoscope lasted for decades: I don't want this. It's going to interfere with my patient visit and bonding. I don't want to learn any sounds.
We are seeing that again. Many physicians are skeptical and worried about AI. Can you address that? Is there a parallel with the stethoscope and these other changes that are occurring?
Rodman: You're right. I did a research study on the death exam, and in the 1890s in the UK, people were still pooh-poohing the stethoscope, four generations of doctors later. In the epistemic approach, knowledge is bounded by certain acceptable ways of accessing information. The stethoscope was a challenge because prior to this, the pathologic anatomy was a whole different way of thinking. Physicians would have said, "Well, I just talked to the patient and then I fit them into a bucket and that's all I need to do."
It was reframing the entire way we thought about disease; now we would have to investigate. People pushed back against that. We're seeing something similar with LLMs, and I have sympathy for those who are skeptical, for a couple of reasons.
One is that there's a lot of hype out there. Martin Shkreli said he had invented an LLM that will replace doctors. I want to reassure everybody that that's not happening anytime soon. The people who want to sell you something are likely to be overhyping. And there have been hyped technologies in the past that have not panned out.
People who are skeptical should try to seriously engage with it. Go to a conference where people are using it. Try it yourself; take 30 minutes to become proficient in using it. At the beginning you will think this changes everything and you'll be super-excited. Then as you use it more, you'll realize that there are uses and limitations, but you'll still understand how useful and potentially transformative the technology is.
Topol: In the August 3 edition of NEJM, you wrote that if we don't shape our own future, powerful technology companies will happily shape it for us. That's pretty important. Can you back it up?
Rodman: Yes. I would draw parallels to the EHR, which was developed in the 1990s. This will shock you, but doctors loved it. Reading about their experiences, they say things like, "It's making me smarter. It's making me take better care of my patients. I'm spending more time with my patients." Doctors used to look forward to using the EHR, like Star Trek — the future. But now, I would bet that many of our listeners are miserable about it.
What happened? Clinicians didn't drive it for their own needs, which was mostly taking care of patients and communicating. It was taken over by large corporate concerns who were more interested in money than usability. The reason I'm so fired up now is that I see this as a transition point. If clinicians and patients and others are interested in driving this technology in the right way, we can have technology that brings us to the sickbay on the Enterprise to spend more time with patients.
I have someone looking over my shoulder, giving me advice when I need it. It strengthens relationships with my patients. But these tools could also be used to extract even more efficiency and money out of a beleaguered healthcare system. Of course they can. They're very powerful. Tech companies aren't evil. They're creating a very powerful tool. I just worry that in our healthcare environment, who is going to drive this forward? Selfishly, I want it to be us and our patients.
Topol: Last week, we heard about the case of a young boy who had symptoms of pain, a rash, growth and development arrest — all sorts of things going wrong in his life. He was seen by 17 different doctors and a dentist over a 3-year period, had extensive workups, and had many tests repeated along the way. After all this, still no diagnosis; every diagnosis he was given proved to be wrong. His mother entered his symptoms into ChatGPT and immediately got the right diagnosis.
Rodman: Tethered cord, right?
Topol: Yes, spina bifida, and he had surgery to untether his cord. Now he is normal — perfectly cured — through ChatGPT. Nothing to do with clinicians; rather, in spite of clinicians. You have thought and written a lot about the clinical community getting up to speed, which hasn't happened yet because all this is happening at a velocity we've never seen before. What about the patient side?
Rodman: This is what excites me. We have open notes now. That's federal law. But those notes are inaccurate. They're full of jargon. They're overwhelming. You can easily imagine, even right now with GPT-4, a custom embedding of a patient's own chart that would allow a patient to talk to their own chart and get a deeper understanding of what's going on, and maybe even see inaccuracies and participate more in their own care. This is controversial, but in the not-so-distant future, LLMs are going to be helping to triage patients in a way that might actually be more patient-friendly because they have infinite patience. They are seen as more empathetic. We haven't tested this, and it needs to be tested, but they may be able to do some things better. They may be able to take a sexual history better than a human because you're talking to a computer. On both sides — involving patients more in their own care, as well as helping clinicians take better care of patients — I see lots of opportunity.
Topol: How do we get medical students and physicians who are in practice up to speed? No one has a background on this; we still don't even fully understand how these new models work. But I think you are right. If we don't do this within the medical community, if we don't take ownership, we're in for some trouble. The University of Texas, San Antonio, just started a combined program for medical students to get an MD degree and an MS degree in AI. At Harvard, they started a PhD in AI and medicine, which is really interesting. Those are the only two programs that have been announced. How are we going to get education going?
Rodman: Our new dean of education, Bernard Chang at Harvard, wrote a piece in JAMA just last week giving his view of what undergraduate medical education should look like in the world of LLMs. It's preparing people to work with machines, not replace them, but to work well together.
This is an active area of debate. I'm just going to acknowledge that the community is split on this. Some people say doctors need to become more proficient with data science; we need to understand how data science works so that we can have two-way communication with data scientists to make things better. I'm not sure about that because if you look at the EHR, that's not a core competency of being a doctor. My general approach has been to think of LLMs as technological tools. We don't need to understand physics to understand how to interpret an MRI. I don't need to understand acoustics to use a stethoscope. Those are my biases as a historian. I don't know that we need doctors to have in-depth knowledge about the inner workings of the technology as much as being proficient in working with it and understanding how it works.
No one knows the answer right now. We're in the very early days. Everyone sees that it's going to be very impactful. Data are starting to come in for some early use cases. But if you were to ask how we should begin in medical education right now, no one knows. The lesson that I take, and this is what I'm trying to do in my residency, is to be open-minded, experiment, study, and evaluate the different things that we're doing and understand that we need to get people proficient in using the technology. That takes practice. That's the academic answer, right? What we need is more study, and that is what I'm trying to do.
Topol: That gets me to another pivotal question. Let's fast-forward 5 years and say, hypothetically, that LLMs are integrated into practice, in the care of patients and in diagnostic and clinical reasoning. It's not done solo. It's done with oversight; humans are always in the loop. So, why do we need to admit brainiacs to medical school, those with the highest SAT or MCAT scores and GPAs through the roof? Why do we need a culture of brainiacs when we have the assistance of machines? What are your thoughts about that?
Rodman: By brainiacs, you mean people who would do very well on standardized tests in the current system, right?
Topol: That's how we select medical students. They don't get an interview. They don't even get to the cut line unless they have excellent GPAs and MCAT scores.
Rodman: I don't think we need that right now. I'm an internist. You're an internist. What do I do as part of my cognitive work? Yes, I do make difficult diagnoses sometimes. But what do I mostly do? I talk to people. I build their trust. I communicate difficult concepts to them. I communicate with ever expanding multidisciplinary teams. I do transitions of care. And yes, I document and put my thoughts on paper. If we were honest about the cognitive work of a doctor, diagnosis is a relatively small part of it. The things that we spend most of our time doing, we get no training on in medical school.
Topol: I want to challenge that because there's really solid documentation that more than 12 million serious diagnostic medical errors are made per year. That's a lot of errors. People are harmed by having the wrong diagnosis.
Rodman: I'm not saying it's not important. I'm saying it's a relatively small part of our cognitive work. It's very important. I quoted that study today. I think over half of the errors were in only 14 diseases, if I recall. We're not messing up lupus nephritis; we're messing up heart disease. It is important, but it's just a small part of what human doctors do.
Medical school focuses on a lot of the other things that make humans good doctors: communication, care transitions, explaining difficult concepts. I'm hopeful that diagnostic errors will decrease dramatically and we'll be better at diagnosis and working with AI. That's a logical conclusion. Would you agree?
Topol: Yes. That was the premise of Deep Medicine, which I wrote 4 or 5 years ago. If we use these AI tools (and we didn't have LLMs) and the gift of time, we will be able to listen to patients rather than interrupting them after 10 seconds. We would actually reestablish the human bond, which has been dreadful over the decades. The type of people who are selected to become physicians might change as well. But when we talk about a big transformation, this might be the biggest one in the past century. That's saying a lot. You are a critical and balanced thinker; you aren't one to support hype. We have a situation where we could get medicine back to emphasizing the trust, the presence, and the empathy of communication. Would you agree that this is our one big shot?
Rodman: I agree with you 100%. The reason that I'm so passionate about this is that if you look at the trajectory of medicine, it has not been heading in that direction. Medicine has become more fragmented, less human, with more errors creeping in. Part of that is a consequence of getting better at treating disease. We've industrialized medicine. We might transition to a period when diseases are defined by these large associations via neural networks that no human being can understand. But it will work. Where does that leave us? What is the professional role of a physician? What does it mean to go to medical school and be a doctor? We don't even know what's coming, what it's going to do to our professional identity. We know something's coming, but we don't know exactly what. And because of the nature of how we founded our knowledge, maybe we can't know. And that's scary and exciting.
Topol: As you are well aware, there's global burnout among clinicians — doctors, nurses, and many other health professionals. I wonder how much of that stems from the inability to provide care because we have so little time with patients. We are spending a lot of time as data clerks, working on preauthorizations and all sorts of menial tasks that are the last thing we want to do with our lives. What are your thoughts about that? Will AI support reduce burnout? Can we start to turn around this crisis or is the crisis truly independent of anything that might be modulated by technology?
Rodman: No, I think you're right. Burnout is multifactorial, but if you look at that time-tracking study by Lena Mamykina published [in 2016], residents spent less than 6 minutes a day with each of their patients, which was less time than they spent walking around or waiting for an elevator. This is not my specialty of research, but that's dehumanizing. Of course, people will burn out if you do that.
If you look at the near-term use cases — what we can do right now — referencing the study we talked about, obviously medicine is highly regulated, which is probably a good thing. But you could probably just have ChatGPT do a lot of our documentation right now, which would give us more time to spend with our patients, even if we froze technology in its place at this very moment. So, I think you're right.
Topol: For listeners who are not up on the latest preprints, the summarization one that you alluded to was very interesting. It suggested that rather than having extended conversations with a chat bot, if you could just summarize everything, you'll get even more accurate answers. That was a very interesting finding because the conversation, as I think you'll agree, is fun. It's fun to talk to what appears to be a fairly astute chat bot, a machine talking back to you. But it gets distracted. It gets off the target of your prompt and objective, so that's a very interesting finding and we'll see if that gets backed up by others.
I really respect your work. I haven't met too many historians who are also futurists. I'm going to look to you to help us find ways to get our medical community up to speed because we aren't there yet. The things that you're working on hopefully are going to help propel us forward. So, thank you, Adam. I appreciate your joining me today.
Follow Medscape on Facebook, X (formerly known as Twitter), Instagram, and YouTube
Medscape © 2023 WebMD, LLC
Any views expressed above are the author's own and do not necessarily reflect the views of WebMD or Medscape.
Cite this: Physicians: Shape Your AI Future or Someone Else Will - Medscape - Oct 02, 2023.
Comments