Transkrip penuh untuk Navigating medical education in the era of generative AI

Morgan Cheatham is a graduate of Brown   00:02:24.480 --> 00:02:29.760 University's Warren Alpert Medical School with  clinical training in genetics at Harvard and is   00:02:29.760 --> 00:02:35.840 a clinical fellow at Boston Children's Hospital.  While Morgan is a bona fide doctor in training,   00:02:35.840 --> 00:02:41.920 he’s also amazingly an influential health  technology strategist. He was recently named   00:02:41.920 --> 00:02:47.520 partner and head of healthcare and life sciences  at Breyer Capital and has led investments in   00:02:47.520 --> 00:02:52.720 several healthcare AI companies that have  eclipsed multibillion-dollar valuations.  00:02:52.720 --> 00:02:57.200 Daniel Chen is finishing his second year as  a medical student at the Kaiser Permanente   00:02:57.200 --> 00:03:02.400 Bernard J. Tyson School of Medicine. He holds  a neuroscience degree from the University of   00:03:02.400 --> 00:03:07.920 Washington and was a research assistant in  the Raskind Lab at the UW School of Medicine,   00:03:07.920 --> 00:03:13.840 working with imaging and genetic data analyses  for biomedical research. Prior to med school,   00:03:13.840 --> 00:03:17.600 Daniel pursued experiences that cultivated  his interest in the application of   00:03:17.600 --> 00:03:23.840 AI in medical practice and education. Daniel and Morgan exemplify the real-world   00:03:23.840 --> 00:03:30.000 future of healthcare, a student entering his third  year of medical school and a fresh medical-school   00:03:30.000 --> 00:03:34.880 graduate who is starting a residency while at  the same time continuing his work on investing   00:03:34.880 --> 00:03:36.960 in healthcare startups. [TRANSITION MUSIC]  00:03:36.960 --> 00:03:48.880 Here is my interview with Morgan Cheatham: LEE: Morgan, thanks for joining. Really,   00:03:48.880 --> 00:03:51.920 really looking forward to this chat. MORGAN CHEATHAM: Peter, it's a privilege   00:03:51.920 --> 00:03:56.400 to be here with you. Thank you. LEE: So are there any other human   00:03:56.400 --> 00:04:04.160 beings who are partners at big-league  venture firms, residents at, you know,   00:04:04.160 --> 00:04:13.120 a Harvard-affiliated medical center, author,  editor for a leading medical journal? I mean, who   00:04:13.120 --> 00:04:18.720 are your … who's your cohort? Who are your peers? CHEATHAM: I love this question. There are so many   00:04:18.720 --> 00:04:23.840 people who I consider peers that I look up to  who have paved this path. And I think what is   00:04:23.840 --> 00:04:29.440 distinct about each of them is they have this  physician-plus orientation. They are multilingual   00:04:29.440 --> 00:04:34.480 in terms of knowing the language of medicine but  having learned other disciplines. And we share   00:04:34.480 --> 00:04:40.960 a common friend, Dr. Zak Kohane, who was among  the first to really show how you can meld two   00:04:40.960 --> 00:04:45.760 worlds as a physician and make

significant  contributions to the intersections thereof.  00:04:45.760 --> 00:04:51.040 I also deeply, in the world of business,  respect physicians like Dr. Krishna Yeshwant   00:04:51.040 --> 00:04:56.960 at Google Ventures, who simultaneously pursued  residency training and built what is now,   00:04:56.960 --> 00:05:00.320 you know, a large and enduring venture firm. So there are plenty of people out there   00:05:00.320 --> 00:05:06.800 who've carved their own path and become these  multilingual beings, and I aspire to be one.  00:05:06.800 --> 00:05:12.400 LEE: So, you know, one thing I've been trying  to explore with people are their origins with   00:05:12.400 --> 00:05:20.080 respect to the technology of AI. And there's  two eras for that. There's AI before ChatGPT   00:05:20.080 --> 00:05:25.600 and before, you know, generative AI really  became a big thing, and then afterwards.  00:05:25.600 --> 00:05:32.320 So let's start first before ChatGPT. You know,  what was your contact? What was, you know,   00:05:32.320 --> 00:05:38.720 your knowledge of AI and machine learning? CHEATHAM: Sure, so my experiences in computer   00:05:38.720 --> 00:05:44.880 science date back to high school. I went to Thomas  Jefferson, which is a high school in Alexandria,   00:05:44.880 --> 00:05:49.200 Virginia, that prides itself on requiring  students to take computer science in their   00:05:49.200 --> 00:05:53.680 first year of high school as kind of a required  torturous experience. [LAUGHTER] And I remember   00:05:53.680 --> 00:05:58.720 that fondly. Our final project was Brick Breaker.  It was actually, I joke, all hard coded. So there   00:05:58.720 --> 00:06:04.240 was nothing intelligent about the Brick Breaker  that we built. But that was my first exposure.  00:06:04.240 --> 00:06:10.160 I was a classics nerd, and I was really interested  in biology and chemistry as a pre-medical student.   00:06:10.160 --> 00:06:15.200 So I really wouldn't intersect with this field  again until I was shadowing at Inova Hospital,   00:06:15.200 --> 00:06:19.840 which was a local hospital near me. And it was  interesting because, at the time—I was shadowing   00:06:19.840 --> 00:06:25.016 in the anesthesia department—they were actually  implementing their first instance of Epic.  00:06:25.016 --> 00:06:26.560 LEE: Mmm. Wow. CHEATHAM: And I remember that   00:06:26.560 --> 00:06:31.840 experience fondly because the entire hospital was  going from analog—they were going from paper-based   00:06:31.840 --> 00:06:37.680 charts—to this new digital system. And I didn't  quite know in that moment what it would mean for   00:06:37.680 --> 00:06:41.440 the field or for my career, but I knew it was  a big deal because a lot of people had a lot of   00:06:41.440 --> 00:06:46.560 emotion around what was going on, and it was in  that experience that I kind of decided to attach   00:06:46.560 --> 00:06:52.720 myself to the intersection of computation  and

medicine. So when I got to undergrad,   00:06:52.720 --> 00:06:58.000 I was a pre-medical student. I was very passionate  about studying the sacred physician-patient   00:06:58.000 --> 00:07:02.480 relationship and everything that had to go on  in that exam room to provide excellent care.  00:07:02.480 --> 00:07:05.280 But there were a few formative experiences: one,   00:07:05.280 --> 00:07:10.160 working at a physician-founded startup that  was using at the time we called it big data,   00:07:10.160 --> 00:07:11.666 if you remember, … LEE: Yup.  00:07:11.666 --> 00:07:17.840 CHEATHAM: … to match the right patient to the  right provider at the right time. And it was in   00:07:17.840 --> 00:07:24.560 that moment that I realized that as a physician,  I could utilize technology to scale that sacred   00:07:24.560 --> 00:07:28.400 one-to-one patient-provider interaction  in nonlinear ways. So that was, kind of,   00:07:28.400 --> 00:07:33.120 the first experience where I saw deployed  systems that were using patient data and   00:07:33.120 --> 00:07:39.840 clinical information in an intelligent format. LEE: Yeah. And so you're a pre-medical student,   00:07:40.480 --> 00:07:48.880 but you have this computer science understanding.  You have an intuition, I guess is the right way to   00:07:48.880 --> 00:07:58.720 say it, that the clinical data becoming digital  is going to be important. So then what happens   00:07:58.720 --> 00:08:03.920 from there to your path to medical school? CHEATHAM: Yeah, so I had a few formative   00:08:03.920 --> 00:08:08.800 research experiences in my undergraduate years.  You know, nothing that ever amounted to a   00:08:08.800 --> 00:08:13.600 significant publication, but I was toying around  with SVMs [support vector machines] for sepsis   00:08:13.600 --> 00:08:15.920 and working with the MIMIC [Medical Information  Mart for Intensive Care] database early days   00:08:15.920 --> 00:08:21.760 and really just trying to understand what it  meant that medical data was becoming digitized.  00:08:21.760 --> 00:08:26.400 And at the same time, again, I was rather  unsatisfied doing that purely in an academic   00:08:26.400 --> 00:08:31.120 context. And I so early craved seeing how  this would roll out in the wild, roll out   00:08:31.120 --> 00:08:35.520 in a clinical setting that I would soon occupy.  And that was really what drove me to work at this   00:08:35.520 --> 00:08:41.360 company called Kyruus [Health] and understand how  these systems, you know, scaled. Obviously, that's   00:08:41.360 --> 00:08:46.080 something with AI that we're now grappling with  in a real way because it looks much different.  00:08:46.080 --> 00:08:48.320 LEE: Right. Yep. CHEATHAM: So the other experience I had,   00:08:48.320 --> 00:08:52.560 which is less relevant to AI, but I did do a  summer in banking. And I mention this because   00:08:52.560 --> 00:08:56.880 what I learned in the experience was … it was  a master class in business. And I learned that   00:08:56.880 --> 00:09:01.760 there was another scaling factor that I should  appreciate as we think about medicine, and that   00:09:01.760 --> 00:09:06.480 was capital and business formation. And that  was also something that could scale nonlinearly.  00:09:06.480 --> 00:09:10.640 So when you married that with technology, it was,  kind of, a natural segue for me before going to   00:09:10.640 --> 00:09:14.880 med school to think about venture capital and  partnering with founders who were going to be   00:09:14.880 --> 00:09:19.200 building these technologies for the long term.  And so that's how I landed on the venture side.  00:09:19.200 --> 00:09:24.960 LEE: And then how long of a break  before you started your medical studies?  00:09:24.960 --> 00:09:28.640 CHEATHAM: It was about four years. Originally,  it was going to be a two-year deferral,   00:09:28.640 --> 00:09:34.320 and the pandemic happened. Our space became quite  active in terms of new companies and investment.   00:09:34.320 --> 00:09:40.640 So it was about four years before I went back. LEE: I see. And so you're in medical school.   00:09:40.640 --> 00:09:44.520 ChatGPT happened while you were  in medical school, is that right?  00:09:44.520 --> 00:09:49.600 CHEATHAM: That's right. That's right. Right before  I was studying for Step 1. So the funny story,   00:09:49.600 --> 00:09:52.240 Peter, that I like to share with folks is … LEE: Yeah.  00:09:52.240 --> 00:09:56.160 CHEATHAM: … I was just embarking on  designing my Step 1 study plan with my   00:09:56.160 --> 00:10:00.880 mentor. And I went to NeurIPS [Conference]  for the first time. And that was in 2022,   00:10:00.880 --> 00:10:08.960 when, of course, ChatGPT was released. And for the remainder of that fall period,   00:10:08.960 --> 00:10:11.816 you know, I should have been studying for these  shelf exams and, you know, getting ready …  00:10:11.816 --> 00:10:12.400 LEE: Yeah. CHEATHAM: … for this   00:10:12.400 --> 00:10:16.800 large board exam. And I was fortunate to partner,  actually, with one of our portfolio company CEOs   00:10:16.800 --> 00:10:24.560 who is a physician—he is an MD/PhD—to work on the  first paper that showed that ChatGPT could pass   00:10:24.560 --> 00:10:26.630 the US Medical Licensing Exam. LEE: Yes.  00:10:26.630 --> 00:10:30.240 CHEATHAM: And that was a riveting experience  for a number of reasons. I joke with folks   00:10:30.240 --> 00:10:36.400 that it was both the best paper I was ever, you  know, a part of and proud to be a coauthor of,   00:10:36.400 --> 00:10:39.520 but also the worst for a lot of  reasons that we could talk about.  00:10:39.520 --> 00:10:44.560 It was the best in terms of canonical metrics  like citations, but the worst in terms of, wow,   00:10:44.560 --> 00:10:47.655 did we spend six months as a field thinking  this was the right benchmark … [LAUGHTER]  00:10:47.655 --> 00:10:49.520 LEE:

Right. CHEATHAM: … for how to assess the performance   00:10:49.520 --> 00:10:52.480 of these models. And I'm so encouraged … LEE: You shouldn't feel bad that way because,   00:10:52.480 --> 00:10:58.160 you know, at that time, I was secretly, you  know, assessing what we now know of as GPT-4   00:10:58.160 --> 00:11:03.680 in that period. And what was the first  thing I tried to do? Step 1 medical exam.  00:11:03.680 --> 00:11:08.480 By the way, just for our listeners who don't  understand about medical education—in the US,   00:11:08.480 --> 00:11:13.360 there's a three-part exam that extends over  a couple of years of medical school. Step 1,   00:11:13.360 --> 00:11:19.360 Step 2, Step 3. And Step 1 and Step 2  in particular are multiple-choice exams.  00:11:19.360 --> 00:11:26.560 And they are very high stakes when you're in  medical school. And you really have to have   00:11:26.560 --> 00:11:32.960 a command of quite a lot of clinical knowledge  to pass these. And it's funny to hear you say   00:11:32.960 --> 00:11:39.680 what you were just sharing there because it  was also the first impulse I had with GPT-4.   00:11:39.680 --> 00:11:44.880 And in retrospect, I feel silly about that. CHEATHAM: I think many of us do, but I've   00:11:44.880 --> 00:11:49.200 been encouraged over the last two years, to your  point, that we really have moved our discourse   00:11:49.200 --> 00:11:55.520 beyond these exams to thinking about more robust  systems for the evaluation of performance, which   00:11:55.520 --> 00:11:59.760 becomes even more interesting as you and I have  spoken about these multi-agent frameworks that we   00:11:59.760 --> 00:12:06.480 are now, you know, compelled to explore further. LEE: Yeah. Well, and even though I know you're a   00:12:06.480 --> 00:12:11.920 little sheepish about it now, I think in the  show notes, we'll link to that paper because   00:12:11.920 --> 00:12:19.040 it really was one of the seminal moments  when we think about AI, AI in medicine.  00:12:19.040 --> 00:12:25.680 And so you're seeing this new technology,  and it's happening at a moment when you   00:12:25.680 --> 00:12:32.640 yourself have to confront taking the  Step 1 exam. So how did that feel?  00:12:32.640 --> 00:12:41.920 CHEATHAM: It was humbling. It was shocking. What I  had worked two years for, grueling over textbooks   00:12:41.920 --> 00:12:48.000 and, you know, flashcards and all of the things  we do as medical students, to see a system emerge   00:12:48.000 --> 00:12:53.016 out of thin air that was arguably going to perform  far better than I ever would, no matter how much …  00:12:53.016 --> 00:12:54.960 LEE: Yeah. CHEATHAM: … I was going to study for that exam,   00:12:55.840 --> 00:13:04.960 it set me back. It forced me to interrogate  what my role in medicine would be. And it   00:13:04.960 --> 00:13:09.680 dramatically changed the specialties  that I considered for myself long term.  00:13:09.680 --> 00:13:14.000 And I hope we talk about, you know, how  I stumbled upon genetics and why I'm so   00:13:14.000 --> 00:13:19.280 excited about that field and its evolution in  this computational landscape. I had to do a   00:13:19.280 --> 00:13:24.480 lot of soul searching to relinquish what  I thought it meant to be a physician and   00:13:24.480 --> 00:13:33.280 how I would adapt in this new environment. LEE: You know, one of the things that we   00:13:33.280 --> 00:13:38.400 wrote in our book, I think it was  in a chapter that I contributed,   00:13:38.400 --> 00:13:48.000 I was imagining that students studying for  Step 1 would be able to do it more actively.  00:13:48.000 --> 00:13:55.200 Or you could even do sort of a pseudo-Step  3 exam by having a conversation. You provide   00:13:55.200 --> 00:14:04.640 the presentation of a patient and then have an  encounter, you know, where the ChatGPT is the   00:14:04.640 --> 00:14:09.840 patient, and then you pretend to be the doctor.  And then in the example that we published,   00:14:09.840 --> 00:14:17.360 then you say, “End of encounter.” And then  you ask ChatGPT for an assessment of things.  00:14:17.360 --> 00:14:24.720 So, you know, maybe it all came too late for  Step 1 for you because you were already very   00:14:24.720 --> 00:14:33.760 focused and, you know, had your own kind of study  framework. But did you have an influence or use of   00:14:33.760 --> 00:14:38.480 this kind of technology for Step 2 and Step 3? CHEATHAM: So even for Step 1,   00:14:38.480 --> 00:14:41.680 I would say, it [ChatGPT], you know, dropped in  November. I took it [Step 1 exam] in the spring,   00:14:41.680 --> 00:14:46.960 so I was able to use it to study. But the  lesson I learned in that moment, Peter,   00:14:46.960 --> 00:14:52.080 was really about the importance of trust with  AI and clinicians or clinicians in training,   00:14:52.080 --> 00:14:55.920 because we all have the same resources that  we use for these board exams, right. UWorld   00:14:55.920 --> 00:15:00.560 is this question bank. It's been around forever.  If you're not using UWorld, like, good luck. And   00:15:00.560 --> 00:15:05.200 so why would you deviate off of a well-trodden  path to study for this really important exam?  00:15:05.200 --> 00:15:10.160 And so I kind of adjunctively used GPT  alongside UWorld to come up with more   00:15:10.160 --> 00:15:14.480 personalized explanations for concepts that I  wasn't understanding, and I found that it was   00:15:14.480 --> 00:15:20.160 pretty good and it was certainly helpful for me. Fortunately, I was, you know, able to pass,   00:15:20.160 --> 00:15:25.600 but I was very intentional about dogfooding  AI when I was a medical student, and part   00:15:25.600 --> 00:15:28.960 of that was because I had been a venture  capitalist, and I'd made investments in   00:15:28.960 --> 00:15:34.080 companies whose products I could actually use. And so, you know, Abridge is a

company in the   00:15:34.080 --> 00:15:36.080 scribing space that you and I have talked about. LEE: Yeah.  00:15:36.080 --> 00:15:41.680 CHEATHAM: I was so fortunate in the early days of  their product to not just be a user but to get to   00:15:41.680 --> 00:15:46.240 bring their product across the hospital. I could  bring the product to the emergency department one   00:15:46.240 --> 00:15:49.760 week, to neurology another week, to the PICU  [pediatric intensive care unit], you know,   00:15:49.760 --> 00:15:54.960 the next week, and assess the relative performance  of, you know, how it handled really complex   00:15:54.960 --> 00:15:55.826 genetics cases … LEE: Yeah.  00:15:55.826 --> 00:16:00.640 CHEATHAM: … versus these very challenging  social situations that you often find   00:16:00.640 --> 00:16:06.240 yourself navigating in primary care. So not  only was I emotional about this technology,   00:16:06.240 --> 00:16:10.160 but I was a voracious adopter in the moment. LEE: Yeah, right. And you had a financial   00:16:10.160 --> 00:16:14.080 interest then on top of that, right? CHEATHAM: I was not paid by Abridge to   00:16:14.080 --> 00:16:16.307 use the product, but, you know, I joke that  the team was probably sick of me. [LAUGHTER]  00:16:16.307 --> 00:16:21.440 LEE: No, no, but you were working for a  venture firm that was invested in these,   00:16:21.440 --> 00:16:28.320 right? So all of these things are wrapped  up together. You know, you're having to get   00:16:28.320 --> 00:16:36.960 licensed as a doctor while doing all of this. So I want to get into that investment and new   00:16:36.960 --> 00:16:41.440 venture stuff there, but let's stick just  for a few more minutes on medical education.   00:16:42.880 --> 00:16:47.680 So I mentioned, you know, what we wrote in the  book, and I remember writing the example, you   00:16:47.680 --> 00:16:54.720 know, of an encounter. Is that at all realistic?  Is anything like that … that was pure speculation   00:16:54.720 --> 00:17:01.120 on our part. What's really happening? And then after we talk about what's   00:17:01.120 --> 00:17:03.920 really happening, what do you  think should happen in medical   00:17:03.920 --> 00:17:11.200 education given the reality of generative AI? CHEATHAM: I've been pleasantly surprised talking   00:17:11.200 --> 00:17:16.880 with my colleagues about AI in clinical settings,  how curious people are and how curious they've   00:17:16.880 --> 00:17:21.920 been over the last two years. I think, oftentimes,  we say, oh, you know, this technology really   00:17:21.920 --> 00:17:27.600 is stratified by age and the younger clinicians  are using it more and the older physicians are   00:17:27.600 --> 00:17:33.040 ignoring it. And, you know, maybe that's true  in some regards, but I've seen, you know, many,   00:17:33.040 --> 00:17:38.080 you know, senior attendings pulling up  Perplexity, GPT, more recently OpenEvidence,   00:17:38.080 --> 00:17:43.840 which has been a really essential tool for me  personally at the point of care, to come up   00:17:43.840 --> 00:17:50.720 with the best decisions for our patients. The general skepticism arises when people   00:17:50.720 --> 00:17:56.480 reflect on their own experience in training  and they think, “Well, I had to learn how to   00:17:56.480 --> 00:17:57.600 do it this way.” LEE: Yeah.  00:17:57.600 --> 00:18:03.760 CHEATHAM: “And therefore, you using an AI scribe  to document this encounter doesn't feel right to   00:18:03.760 --> 00:18:09.920 me because I didn't get to do that.” And I did  face some of those critiques or criticisms, where   00:18:09.920 --> 00:18:14.800 you need to learn how to do it the old-school  way first and then you can use an AI scribe.  00:18:14.800 --> 00:18:20.560 And I haven't yet seen—maybe even taking a step  back—I haven't seen a lot of integration of AI   00:18:20.560 --> 00:18:23.257 into the core medical curriculum, period. LEE: Yeah.  00:18:23.257 --> 00:18:25.040 CHEATHAM: And, as you know, if  you want to add something to   00:18:25.040 --> 00:18:31.095 medical school curriculum, you can get in a  long line of people who also want to do that.  00:18:31.095 --> 00:18:32.160 LEE: Yes. Yeah. CHEATHAM: But it is urgent   00:18:32.160 --> 00:18:37.280 that our medical schools do create formalized  required trainings for this technology because   00:18:37.280 --> 00:18:38.520 people are already using it. LEE: Yes.  00:18:38.520 --> 00:18:45.200 CHEATHAM: I think what we will need to determine  is how much of the old way do people need to learn   00:18:45.200 --> 00:18:50.160 in order to earn the right to use AI at the point  of care and how much of that old understanding,   00:18:50.160 --> 00:18:56.320 that prior experience, is essential to be able to  assess the performance of these tools and whether   00:18:56.320 --> 00:19:00.960 or not they are having the intended outcome. I kind of joke it's like learning cursive, right?  00:19:00.960 --> 00:19:01.920 LEE: Yes. CHEATHAM: I'm   00:19:01.920 --> 00:19:04.800 old enough to have had to learn cursive.  I don't think people really have to learn   00:19:04.800 --> 00:19:07.920 it these days. When do I use it?  Well, when I'm signing something.   00:19:07.920 --> 00:19:09.760 I don't even really sign checks anymore, but ... LEE: Well … the example I've used, which you've   00:19:09.760 --> 00:19:15.811 heard, is, I'm sure you were still taught the  technique of manual palpation, even though …  00:19:15.811 --> 00:19:17.120 CHEATHAM: Of course. LEE: … you have access to   00:19:17.120 --> 00:19:22.160 technologies like ultrasound. And in fact,  you would use ultrasound in many cases.  00:19:23.280 --> 00:19:29.200 And so I need to pin you down. What  is your opinion on these things? Do   00:19:29.200 --> 00:19:36.720 you need to be trained in the old ways? CHEATHAM: When it comes to understanding   00:19:36.720 --> 00:19:42.800

the architecture of the medical note, I believe  it is important for clinicians in training to   00:19:42.800 --> 00:19:48.320 know how that information is generated, how  it's characterized, and how to go from a very   00:19:48.320 --> 00:19:54.480 broad-reaching conversation to a distilled  clinical document that serves many functions.  00:19:54.480 --> 00:19:58.400 Does that mean that you should be forced  to practice without an AI scribe for the   00:19:58.400 --> 00:20:02.240 entirety of your medical education? No.  And I think that as you are learning the   00:20:02.240 --> 00:20:06.240 architecture of that document, you should  be handed an AI scribe and you should be   00:20:06.240 --> 00:20:11.920 empowered to have visits with patients both in  an analog setting—where you are transcribing and   00:20:11.920 --> 00:20:16.880 generating that note—and soon thereafter,  I'm talking in a matter of weeks, working   00:20:16.880 --> 00:20:23.120 with an AI scribe. That's my personal belief. LEE: Yeah, yeah. So you're going to ... well,   00:20:23.120 --> 00:20:28.550 first off, congratulations on landing a  residency at Boston Children's [Hospital].  00:20:28.550 --> 00:20:31.920 CHEATHAM: Thank you, Peter. LEE: I understand there were only two   00:20:31.920 --> 00:20:40.160 people selected for this and super competitive.  You know, with that perspective, you know,   00:20:40.160 --> 00:20:50.240 how do you see your future in medicine, just given  everything that's happening with AI right now?  00:20:50.240 --> 00:20:58.400 And are there things that you would urge,  let's say, the dean of the Brown Medical School   00:20:59.200 --> 00:21:05.840 to consider or to change? Or  maybe not the dean of Brown but   00:21:06.480 --> 00:21:11.840 the head of the LCME [Liaison Committee on Medical  Education], the accrediting body for US medical   00:21:11.840 --> 00:21:19.440 schools. What in your mind needs to change? CHEATHAM: Sure. I'll answer your first question   00:21:19.440 --> 00:21:21.520 first and then talk about the future. LEE: Yeah.  00:21:21.520 --> 00:21:26.320 CHEATHAM: For me personally, I fell into  the field of genomics. And so my training   00:21:26.320 --> 00:21:31.440 program will cover both pediatrics as  well as clinical genetics and genomics.  00:21:31.440 --> 00:21:36.480 And I alluded to this earlier, but one of the  reasons I'm so excited to join the field is   00:21:36.480 --> 00:21:41.760 because I really feel like the field of genetics  not only is focused on a very underserved patient   00:21:41.760 --> 00:21:46.960 population, but not in how we typically think of  underserved. I'm talking about underserved as in   00:21:46.960 --> 00:21:51.840 patients who don't always have answers.  Patients for whom the current guidelines   00:21:51.840 --> 00:21:58.960 don't offer information or comfort or support. Those are patients that are extremely underserved.   00:21:58.960 --> 00:22:04.080 And I think in this moment of AI, there's a  unique opportunity to utilize the computational   00:22:04.080 --> 00:22:10.000 systems that we now have access to, to provide  these answers more precisely, more quickly.  00:22:10.720 --> 00:22:14.480 And so I'm excited to marry those two fields.  And genetics has long been a field that has   00:22:14.480 --> 00:22:18.640 adopted technology. We just think about the  foundational technology of genomic sequencing   00:22:18.640 --> 00:22:23.040 and variant interpretation. And so it's a kind  of natural evolution of the field, I believe,   00:22:23.040 --> 00:22:30.320 to integrate AI and specifically generative AI. If I were speaking directly to the LCME,   00:22:30.320 --> 00:22:35.600 I mean, I would just have to encourage the  organization, as well as medical societies   00:22:35.600 --> 00:22:39.920 who partner with attending physicians  across specialties, to lean in here.  00:22:39.920 --> 00:22:44.960 When I think about prior revolutions in technology  and medicine, physicians were not always at the   00:22:44.960 --> 00:22:50.800 helm. We have a unique opportunity now, and you  talk about companies like Abridge in the AI space,   00:22:50.800 --> 00:22:55.760 companies like Viz.ai, Cleerly—I mean, I  could go on: Iterative Health ... I could   00:22:55.760 --> 00:23:02.560 list 20 organizations that are bringing AI to  the point of care that are founded by physicians.  00:23:02.560 --> 00:23:07.680 This is our moment to have a seat at the  table and to shape not only the discourse   00:23:07.680 --> 00:23:12.640 but the deployment. And the unique lens,  of course, that a physician brings is that   00:23:12.640 --> 00:23:18.720 of prioritizing the patient, and with AI  and this time around, we have to do that.  00:23:18.720 --> 00:23:25.680 LEE: So LCME for our listeners is, I  think it stands for the Liaison Committee   00:23:25.680 --> 00:23:30.960 on Medical Education. It's basically the  accrediting body for US medical schools,   00:23:30.960 --> 00:23:36.320 and it's very high stakes. It's very, very  rigorous, which I think is a good thing,   00:23:36.320 --> 00:23:43.440 but it's also a bit of a straitjacket. So if you are on the LCME,   00:23:44.240 --> 00:23:53.280 are there specific new curricular elements  that you would demand that LCME, you know,   00:23:53.280 --> 00:23:59.200 add to its accreditation standards? CHEATHAM: We need to unbundle the different   00:23:59.200 --> 00:24:03.680 components of the medical appointment  and think about the different functions   00:24:03.680 --> 00:24:09.200 of a human clinician to answer that question. There are a couple of areas that are top of mind   00:24:09.200 --> 00:24:15.120 for me, the first being medical search. There are  large organizations and healthcare incumbents that   00:24:15.120 --> 00:24:20.480 have been around for many decades, companies like  UpToDate or even, you know, the guidelines that   00:24:20.480 -->

00:24:26.720 are housed by our medical societies, that need to  respond to the information demands of clinicians   00:24:26.720 --> 00:24:31.520 at the point of care in a new way with AI. And so I would love to see our medical   00:24:31.520 --> 00:24:37.360 institutions teaching more students how to use  AI for medical search problems at the point of   00:24:37.360 --> 00:24:42.960 care. How to not only, you know, from a prompt  perspective, ask questions about patients in   00:24:43.600 --> 00:24:48.400 a high-efficacy way, but also to interpret  the outputs of these systems to inform   00:24:48.400 --> 00:24:53.280 downstream clinical decision-making. People are already adopting, as you know,   00:24:53.280 --> 00:24:57.840 GPT, OpenEvidence, Perplexity, all of  these tools to make these decisions now.  00:24:57.840 --> 00:25:03.120 And so by not—again, it's a moral imperative of  the LCME—by not having curriculum and support   00:25:03.120 --> 00:25:07.920 for clinicians doing that, we run the risk  of folks not utilizing these tools properly   00:25:07.920 --> 00:25:12.320 but also to their greatest potential. LEE: Yeah, then, but zooming forward   00:25:12.320 --> 00:25:18.160 then, what about board certification? CHEATHAM: Board certification today is already   00:25:18.160 --> 00:25:22.640 transitioning to an open-book format for many  specialties, is my understanding. And in talking   00:25:22.640 --> 00:25:26.640 to some of my fellow geneticists, who, you know,  that's a pretty challenging board exam in clinical   00:25:26.640 --> 00:25:32.800 genetics or biochemical genetics. They are  using OpenEvidence during those open-book exams.  00:25:32.800 --> 00:25:39.120 So what I would like to see us do is move from  a system of rote memorization and regurgitation   00:25:39.120 --> 00:25:48.000 of fact to an assessment framework that is  adaptive, is responsive, and assesses for   00:25:48.000 --> 00:25:52.880 your ability to use the tools that we now have  at our disposal to make sound clinical decisions.  00:25:52.880 --> 00:26:00.640 LEE: Yeah. We've heard from Sara Murray, you know,  that when she's doing her rounds, she consults   00:26:00.640 --> 00:26:05.680 routinely with ChatGPT. And that was something we  also predicted, especially Carey Goldberg in our   00:26:05.680 --> 00:26:11.920 book, you know, wrote this fictional account. Is that the primary real-world use of AI?   00:26:11.920 --> 00:26:17.760 Not only by clinicians, but also by medical  students ... are medical students, you know,   00:26:17.760 --> 00:26:24.560 engaged with ChatGPT or, you know, similar? CHEATHAM: Absolutely. I've listed some of the   00:26:24.560 --> 00:26:30.000 tools. I think there, in general, Peter, there is  this new clinician stack that is emerging of these   00:26:30.000 --> 00:26:34.960 tools that people are trying, and I think the  cycles of iteration are quick, right. Some folks   00:26:34.960 --> 00:26:38.960 are using Claude [Claude.ai] one week, and they're  trying Perplexity, or they're trying OpenEvidence,   00:26:38.960 --> 00:26:43.120 they're trying GPT for a different task. There's this kind of moment in medicine that every   00:26:43.120 --> 00:26:48.160 clinician experiences where you're on rounds, and  there's that very senior attending. And you've   00:26:48.160 --> 00:26:52.000 just presented a patient to them, and you think  you did an elegant job, and you've summarized all   00:26:52.000 --> 00:26:56.640 the information, and you really feel good about  your differential, and they ask you, like, the one   00:26:56.640 --> 00:27:01.920 question you didn't think to address. [LAUGHTER] And I'll tell you, some of the funniest moments   00:27:01.920 --> 00:27:06.320 I've had using AI in the hospital has  been, and let me take a step back,   00:27:06.320 --> 00:27:10.480 that process of an attending physician  interrogating a medical student is   00:27:10.480 --> 00:27:15.600 called “pimping,” for lack of a better phrase. And some of the funniest use cases I've had for   00:27:15.600 --> 00:27:21.680 AI in that setting is actually using OpenEvidence  or GPT as defense against pimping. [LAUGHTER] So   00:27:21.680 --> 00:27:25.840 quickly while they're asking me the question,  I put it in, and I'm actually able to answer it   00:27:25.840 --> 00:27:29.520 right away. So it's been effective for that.  But I would say, you know, [in] the halls of   00:27:29.520 --> 00:27:34.800 most of the hospitals where I've trained,  I'm seeing this technology in the wild.  00:27:34.800 --> 00:27:42.480 LEE: So now you're so tech-forward,  but that off-label use of AI, we also,   00:27:42.480 --> 00:27:48.880 when we wrote our book, we weren't sure that  at least top health systems would tolerate   00:27:48.880 --> 00:27:56.080 this. Do you have an opinion about this? Should  these things be better regulated or controlled   00:27:56.080 --> 00:28:01.040 by the CIOs of Boston Children's? CHEATHAM: I'm a big believer that   00:28:01.040 --> 00:28:06.640 transparency encourages good behaviors. And so the first time I actually tried to   00:28:06.640 --> 00:28:10.720 use ChatGPT in a clinical setting, it was at a  hospital in Rhode Island. I will not name which   00:28:10.720 --> 00:28:15.280 hospital. But the site was actually blocked.  I wasn't able to access it from a desktop.   00:28:15.280 --> 00:28:18.800 That was the hospital's first response to  this technology was, let's make sure none   00:28:18.800 --> 00:28:22.960 of our clinicians can access it. It has so much  potential for medicine. The irony of that today.  00:28:22.960 --> 00:28:27.680 And it's since, you know, become unblocked. But I  was able to use it on my phone. So, to your point,   00:28:27.680 --> 00:28:31.200 if there's a will, there's a way.  And we will utilize this technology   00:28:31.200 --> 00:28:37.840 if we are seeing perceived value. LEE: So, yeah, no,

absolutely. So now,   00:28:38.800 --> 00:28:46.320 you know, in some discussions, one  superpower that seems to be common across   00:28:46.320 --> 00:28:55.200 people who are really leading the charge here is  they seem to be very good readers and students.  00:28:55.200 --> 00:28:59.680 And I understand you also as  a voracious reader. In fact,   00:28:59.680 --> 00:29:08.160 you're even on an editorial team for a major  medical journal. To what extent does that help?  00:29:08.160 --> 00:29:11.760 And then from your vantage point at  New England Journal of Medicine AI—and   00:29:11.760 --> 00:29:18.960 I'll have a chance to ask Zak Kohane as the  editor in chief the same question—you know,   00:29:18.960 --> 00:29:24.720 what's your assessment as you reflect over the  last two years for the submitted manuscripts?   00:29:24.720 --> 00:29:33.520 Are you overall surprised at what you're seeing?  Disappointed? Any kind of notable hits or misses,   00:29:33.520 --> 00:29:39.360 just in the steady stream of research papers that  are getting submitted to that leading journal?  00:29:39.360 --> 00:29:45.680 CHEATHAM: I would say overall, the field  is becoming more expansive in the kinds   00:29:45.680 --> 00:29:49.600 of questions that people are asking. Again, when we started, it was this   00:29:49.600 --> 00:29:54.960 very myopic approach of: “Can we pass these  medical licensing exams? Can we benchmark   00:29:54.960 --> 00:29:59.840 this technology to how we benchmark our human  clinicians?” I think that's a trap. Some folks   00:29:59.840 --> 00:30:04.000 call this the Turing Trap, right, of let's just  compare everything to what a human is capable of.  00:30:04.000 --> 00:30:07.920 Instead of interrogating what is the  unique, as you all talk about in the book,   00:30:07.920 --> 00:30:12.160 what are the unique attributes of this  new substrate for computation and what   00:30:12.160 --> 00:30:16.640 new behaviors emerge from it, whether that's  from a workflow perspective in the back office,   00:30:16.640 --> 00:30:21.680 or—as I'm personally more passionate and as we're  seeing more people focus on in the literature—what   00:30:21.680 --> 00:30:25.760 are the diagnostic capabilities, right. I love Eric Topol's framework for   00:30:25.760 --> 00:30:29.120 “machine eyes,” right, as this notion  of like, yes, we as humans have eyes,   00:30:29.120 --> 00:30:34.160 and we have looked at medical images for many,  many decades, but these machines can take a   00:30:34.160 --> 00:30:39.520 different approach to a retinal image, right. It's not just what you can diagnose in terms of   00:30:39.520 --> 00:30:45.200 an ophthalmological disease but maybe a neurologic  disease or, you know, maybe liver disease, right.  00:30:45.200 --> 00:30:47.600 So I think the literature is,  in general, moving to this   00:30:47.600 --> 00:30:54.160 place of expansion, and I'm excited by that. LEE: Yeah, I kind of have referred to that as   00:30:54.160 --> 00:31:01.760 triangulation. You know, one of the things I think  a trap that specialists in medicine can fall into,   00:31:01.760 --> 00:31:09.760 like a cardiologist will see everything in  terms of the cardiac system. And … whereas   00:31:09.760 --> 00:31:15.600 a nephrologist will see things in a certain lens. And one of the things that you oftentimes see in   00:31:15.600 --> 00:31:24.400 the responses from a large language model is that  more expansive view. At the same time, you know,   00:31:24.400 --> 00:31:33.840 I wonder … we have medical specialties for good  reason. And, you know, at times I do wonder, you   00:31:33.840 --> 00:31:40.000 know, if there can be confusion that builds up. CHEATHAM: This is an under-discussed   00:31:40.000 --> 00:31:45.440 area of AI—AI collapses medical  specialties onto themselves, right.  00:31:45.440 --> 00:31:49.200 You have the canonical example of the  cardiologist, you know, arguing that,   00:31:49.200 --> 00:31:53.120 you know, we should diuresis and maybe  the nephrologist arguing that we should,   00:31:53.120 --> 00:32:00.800 you know, protect the kidneys. And how do two  disciplines disagree on what is right for the   00:32:00.800 --> 00:32:05.680 patient when in theory, there is an objective  best answer given that patient's clinical status?  00:32:07.120 --> 00:32:10.880 My understanding is that the emergence of  medical specialties was a function of the   00:32:10.880 --> 00:32:16.960 cognitive overload of medicine in general and how  difficult it was to keep all of the specifics of   00:32:16.960 --> 00:32:20.240 a given specialty in the mind. Of course,  general practitioners are tasked with doing   00:32:20.240 --> 00:32:23.600 this at some level, but they're also tasked  with knowing when they've reached their limit   00:32:23.600 --> 00:32:28.880 and when they need to refer to a specialist. So I'm interested in this question of whether   00:32:28.880 --> 00:32:33.360 medical specialties themselves need to evolve. And if we look back in the history of medical   00:32:33.360 --> 00:32:39.360 technology, there are many times where a new  technology forced a medical specialty to evolve,   00:32:39.360 --> 00:32:43.520 whether it was certain diagnostic tools that  have been introduced or, as we're seeing now with   00:32:43.520 --> 00:32:47.266 GLP-1s, the entire cardiometabolic field … LEE: Right.  00:32:47.266 --> 00:32:52.160 CHEATHAM: … is having to really reimagine  itself with these new tools. So I think   00:32:52.160 --> 00:32:57.280 AI will look very similar, and we should  not hold on to this notion of classical   00:32:57.280 --> 00:33:05.120 medical specialties simply out of convention. LEE: Right. All right. So now you're starting your   00:33:05.120 --> 00:33:12.960 residency. You're, you know, basically leading a  charge in health and life sciences for a leading   00:33:12.960 --> 00:33:20.320 venture firm. I'd like you

to predict what  the world of healthcare is going to look like,   00:33:20.320 --> 00:33:24.000 you know, two years from now, five years from  now, 10 years from now. And to frame that,   00:33:24.000 --> 00:33:30.080 to make it a little more specific, you know,  what do you think will be possible that you,   00:33:30.080 --> 00:33:34.960 as a doctor and an investor, will be able to  do two years from now, five years from now,   00:33:34.960 --> 00:33:40.720 10 years from now that you can't do today? CHEATHAM: Two years from now, I'm optimistic   00:33:40.720 --> 00:33:48.720 we'll have greater adoption of AI by clinicians,  both for back-office use cases. So whether that's   00:33:48.720 --> 00:33:53.600 the scribe and the generation of the note  for billing purposes, but also now thinking   00:33:53.600 --> 00:33:58.000 about that for patient-facing applications. We're already doing this with drafting of notes.   00:33:58.000 --> 00:34:04.080 I think we'll see greater proliferation of those  more obvious use cases over the next two years.   00:34:04.080 --> 00:34:08.640 And hopefully we're seeing that across hospital  systems, not just large well-funded academics,   00:34:08.640 --> 00:34:14.320 but really reaching our community hospitals, our  rural hospitals, our under-resourced settings.  00:34:14.320 --> 00:34:20.560 I think hopefully we'll see greater conversion.  Right now, we have this challenge of “pilotitis,”   00:34:20.560 --> 00:34:24.560 right. A lot of people are trying things,  but the data shows that only one in three   00:34:24.560 --> 00:34:28.480 pilots are really converting to production use. So  hopefully we'll kind of move things forward that   00:34:28.480 --> 00:34:33.600 are working and pare back on those that are not. We will not solve the problem of payment models in   00:34:33.600 --> 00:34:38.560 the next two years. That is a prediction I have. Over the next five years, I suspect that, with   00:34:38.560 --> 00:34:45.120 the help of regulators, we will identify better  payment mechanisms to support the adoption of   00:34:45.120 --> 00:34:50.640 AI because it cannot and will not sustain itself  simply by asking health systems and hospitals to   00:34:50.640 --> 00:34:55.840 pay for it. That is not a scalable solution. LEE: Yes. Right. Yep. In fact, I think there   00:34:55.840 --> 00:35:02.480 have to be new revenue-positive  incentives if providers are asked   00:35:02.480 --> 00:35:07.760 to do more in the adoption of technology. CHEATHAM: Absolutely. But as we appreciate,   00:35:07.760 --> 00:35:12.240 some of the most promising applications  of AI have nothing to do with revenue.   00:35:12.240 --> 00:35:18.320 It might simply be providing a diagnosis to  somebody, you know, for whom that might drive   00:35:18.320 --> 00:35:21.840 additional intervention, but may also not. And we have to be OK with that because   00:35:21.840 --> 00:35:26.080 that's the right thing to do. It's our moral  imperative as clinicians to implement this   00:35:26.080 --> 00:35:30.480 where it provides value to the patient. Over the next 10 years, what I—again,   00:35:30.480 --> 00:35:38.080 being a techno-optimist—am hopeful we start to  see is a dissolving of the barrier that exists   00:35:38.080 --> 00:35:44.880 between care delivery and biomedical discovery. This is the vision of the learning health system   00:35:44.880 --> 00:35:51.040 that was written over 10 years ago, and we  have not realized it in practice. I'm a big   00:35:51.040 --> 00:35:58.320 proponent of ensuring that every single patient  that enters our healthcare system not only   00:35:58.320 --> 00:36:05.920 receives the best care, but that we learn from the  experiences of that individual to help the next.  00:36:05.920 --> 00:36:11.680 And in our current system, that is not how it  works. But, with AI, that now becomes possible.  00:36:11.680 --> 00:36:17.520 LEE: Well, I think connecting healthcare  experiences to medical discovery—I think   00:36:17.520 --> 00:36:23.200 that that is really such a great vision for  the future. And I do agree [that] AI really   00:36:23.200 --> 00:36:29.360 gives us real hope that we can make it true. Morgan, I think we could talk for a few hours   00:36:29.360 --> 00:36:36.000 more. It's just incredible what you're up  to nowadays. Thank you so much for this   00:36:36.000 --> 00:36:40.960 conversation. I've learned a lot talking to you. CHEATHAM: Peter, thank you so much for your time.   00:36:40.960 --> 00:36:56.857 I will be clutching my signed copy of The AI  Revolution in Medicine for many years to come.  00:36:56.857 --> 00:36:58.960 [TRANSITION MUSIC] LEE: Morgan obviously is   00:36:58.960 --> 00:37:05.680 not an ordinary med school graduate. In previous  episodes, one of the things we've seen is that   00:37:05.680 --> 00:37:12.960 people on the leading edge of real-world AI in  medicine oftentimes are both practicing doctors   00:37:12.960 --> 00:37:19.680 as well as technology developers. Morgan is  another example of this type of polymath,   00:37:19.680 --> 00:37:26.320 being both a med student and a venture capitalist. One thing that struck me about Morgan is he's   00:37:26.320 --> 00:37:34.080 just incredibly hands-on. He goes out, finds  leading-edge tools and technologies, and often   00:37:34.800 --> 00:37:39.120 these things, even though they're experimental,  he takes them into his education and into his   00:37:39.120 --> 00:37:44.720 clinical experiences. I think this highlights a  potentially important point for medical schools,   00:37:44.720 --> 00:37:50.960 and that is, it might be incredibly important  to provide the support—and, let's be serious,   00:37:50.960 --> 00:37:57.040 the permission—to students to access and  use new tools and technologies. Indeed,   00:37:57.040 --> 00:38:03.360 the insight for me when I interact with Morgan is  that in these early days of AI in medicine,

there   00:38:03.360 --> 00:38:12.160 is no substitute for hands-on experimentation, and  that is likely best done while in medical school.  00:38:12.160 --> 00:38:24.227 Here's my interview with Daniel Chen: LEE: Daniel, it's great to have you here.  00:38:24.227 --> 00:38:28.560 DANIEL CHEN: Yeah, it's a pleasure being here. LEE: Well, you know, I normally get started in   00:38:28.560 --> 00:38:33.360 these conversations by asking, you know, how do  you explain to your mother what you do all day?   00:38:33.360 --> 00:38:39.920 And the reason that that's a good question is a  lot of the people we have on this podcast have   00:38:39.920 --> 00:38:49.840 fancy titles and unusual jobs, but I'm guessing  that your mother would have already a preconceived   00:38:49.840 --> 00:38:55.040 notion of what a medical student does. So I'd  like to twist the question around a little   00:38:55.040 --> 00:39:05.040 bit for you and ask, what does your mother not  realize about how you spend your days at school?  00:39:05.040 --> 00:39:09.040 Or does she get it all right? [LAUGHS] CHEN: Oh, no, she is very observant. I'll say   00:39:09.040 --> 00:39:15.440 that off the bat. But I think something that she  might not realize is the amount of efforts spent,   00:39:15.440 --> 00:39:20.560 kind of, outside the classroom or outside  the hospital. You know, she's always,   00:39:20.560 --> 00:39:25.680 like, saying you have such long days in the  hospital. You're there so early in the morning.  00:39:25.680 --> 00:39:29.600 But what she doesn't realize is that maybe when I  come back from the hospital, it's not just like,   00:39:29.600 --> 00:39:34.640 oh, I'm done for the day. Let's wind down, go  to bed. But it's more like, OK, I have some more   00:39:34.640 --> 00:39:40.080 practice questions I need to get through; I didn't  get through my studying. Let me write on, like   00:39:40.080 --> 00:39:44.320 wrap up this research project I'm working on, get  that back to the PI [principal investigator]. It's   00:39:44.320 --> 00:39:48.800 never ending to a certain extent. Those  are some things she doesn't realize.  00:39:48.800 --> 00:39:53.760 LEE: Yeah, I think, you know, all the time  studying, I think, is something that people   00:39:53.760 --> 00:40:01.440 expect of second-year medical students. And even  nowadays at the top medical schools like this one,   00:40:01.440 --> 00:40:06.480 being involved in research is also expected. I think one thing that is a little unusual is   00:40:06.480 --> 00:40:13.800 that you are actually in clinic, as well, as a  second-year student. How has that gone for you?  00:40:13.800 --> 00:40:19.600 CHEN: Yeah, I mean, it's definitely interesting.  I would say I spend my time, especially this year,   00:40:19.600 --> 00:40:22.640 it's kind of three things. There's the  preclinical stuff I'm doing. So that's   00:40:22.640 --> 00:40:26.480 your classic, you know, you're learning  from the books, though I don't feel like   00:40:26.480 --> 00:40:31.600 many of us do have textbooks anymore. [LAUGHTER] There's the clinical aspect, which you mentioned,   00:40:31.600 --> 00:40:35.760 which is we have an interesting model,  longitudinal integrated clerkships. We can   00:40:35.760 --> 00:40:39.680 talk about that. And the last component is the  research aspect, right. The extracurriculars.  00:40:40.560 --> 00:40:46.080 But I think starting out as a second year and  doing your rotations, probably early on in,   00:40:46.080 --> 00:40:53.040 kind of, the clinical medical education, has been  really interesting, especially with our format,   00:40:53.040 --> 00:40:56.960 because typically med students have two  years to read up on all the material and,   00:40:56.960 --> 00:41:02.240 like, get some foundational knowledge. With us,  it's a bit more, we have maybe one year under our   00:41:02.240 --> 00:41:07.120 belt before we're thrown into like, OK, go talk to  this patient; they have ankle pain, right. But we   00:41:07.120 --> 00:41:11.200 might have not even started talking about ankle  pain in class, right. Well, where do I begin?  00:41:11.200 --> 00:41:16.800 So I think starting out, it's kind of, like, you  know, the classic drinking from a fire hydrant.   00:41:16.800 --> 00:41:20.320 But you also, kind of, have that embarrassment  of you're talking to the patient like, I have   00:41:20.320 --> 00:41:25.200 no clue what's happening [LAUGHTER] or you might  have … my differentials all over the place, right.  00:41:25.200 --> 00:41:29.440 But I think the beauty of the longitudinal  aspect is that now that we're, like,   00:41:29.440 --> 00:41:34.960 in our last trimester, everything's kind of  coming together. Like, OK, I can start to see,   00:41:34.960 --> 00:41:38.720 you know, here's what you're telling me.  Here's what the physical exam findings are.   00:41:38.720 --> 00:41:42.240 I'm starting to form a differential. Like,  OK, I think these are the top three things.  00:41:42.240 --> 00:41:46.320 But in addition to that, I think these are the  next steps you should take so we can really focus   00:41:46.320 --> 00:41:53.040 and hone in on what exact diagnosis this might be. LEE: All right. So, of course, what we're   00:41:53.040 --> 00:41:58.400 trying to get into is about AI. And, you know, the funny thing about   00:41:58.400 --> 00:42:05.600 AI and the book that Carey, Zak, and I wrote  is we actually didn't think too much about   00:42:05.600 --> 00:42:11.360 medical education, although we did have some  parts of our book where we, well, first off,   00:42:11.360 --> 00:42:17.680 we made the guess that medical students would find  AI to be useful. And we even had some examples,   00:42:17.680 --> 00:42:25.360 you know, where, you know, you would have  a vignette of a mythical patient, and you   00:42:25.360 --> 00:42:31.360 would ask the AI to pretend to be that patient. And then you would have an

interaction and have   00:42:31.360 --> 00:42:37.440 to have an encounter. And so I want to delve into  whether any of that is happening. How real it is.   00:42:37.440 --> 00:42:44.960 But before we do that, let's get into first off,  your own personal contact with AI. So let me start   00:42:44.960 --> 00:42:51.560 with a very simple question. Do you ever use  generative AI systems like ChatGPT or similar?  00:42:51.560 --> 00:42:55.280 CHEN: All the time, if not every day. LEE: [LAUGHS] Every day,   00:42:55.280 --> 00:43:00.640 OK. And when did that start? CHEN: I think when it first launched   00:43:00.640 --> 00:43:07.520 with GPT-3.5, I was, you know, curious. All my  friends work in tech. You know, they're either   00:43:07.520 --> 00:43:12.320 software engineers, PMs. They're like, “Hey,  Daniel, take a look at this,” and at first,   00:43:12.320 --> 00:43:16.640 I thought it was just more of a glorified search  engine. You know, I was actually looking back.  00:43:16.640 --> 00:43:21.920 My first question to ChatGPT was, what was the  weather going to be like the next week, you know?   00:43:21.920 --> 00:43:26.474 Something very, like, something easily you could  have looked up on Google or your phone app, right.  00:43:26.474 --> 00:43:30.720 I was like, oh, this is pretty cool. But  then, kind of, fast-forwarding to, I think,   00:43:30.720 --> 00:43:37.120 the first instance I was using it in med school.  I think the first, like, thing that really helped   00:43:37.120 --> 00:43:43.920 me was actually a coding problem. It was for  a research project. I was trying to use SQL.  00:43:43.920 --> 00:43:48.400 Obviously, I've never taken a SQL class in my  life. So I asked Chat like, “Hey, can you write   00:43:48.400 --> 00:43:53.760 me this code to maybe morph two columns together,”  right? Something that might have taken me hours to   00:43:53.760 --> 00:43:58.080 maybe Google on YouTube or like try to read some  documentation which just goes through my head.  00:43:58.080 --> 00:44:03.200 But ChatGPT was able to, you know,  not only produce the code, but, like,   00:44:03.200 --> 00:44:06.640 walk me through like, OK, you're going  to launch SQL. You're going to click on   00:44:06.640 --> 00:44:11.040 this menu, [LAUGHTER] put the code in here, make  sure your file names are correct. And it worked.  00:44:11.040 --> 00:44:17.840 So it's been a very powerful tool in that way in  terms of, like, giving me expertise in something   00:44:17.840 --> 00:44:23.920 that maybe I traditionally had no training in. LEE: And so while you're doing this, I assume   00:44:23.920 --> 00:44:31.280 you had fellow students, friends, and others. And  so what were you observing about their contact   00:44:31.280 --> 00:44:39.280 with AI? I assume you weren't alone in this. CHEN: Yeah, yeah, I think, ... I'm not too sure   00:44:39.280 --> 00:44:43.040 in terms of what they were doing when it first  came out, but I think if we were

talking about   00:44:43.040 --> 00:44:48.240 present day, um, a lot of it's kind of really  spot on to what you guys talked about in the book.  00:44:48.240 --> 00:44:54.880 Um, I think the idea around this personal tutor,  personal mentor, is something that we're seeing a   00:44:54.880 --> 00:45:00.160 lot. Even if we're having in-class discussions,  the lecturer might be saying something, right.   00:45:00.160 --> 00:45:06.800 And then I might be or I see a friend in ChatGPT  or some other model looking up a question.  00:45:07.520 --> 00:45:12.640 And you guys talked about, you know, how it can,  like, explain a concept at different levels,   00:45:12.640 --> 00:45:16.960 right. But honestly, sometimes if there's a  complex topic, I ask ChatGPT, like, can you   00:45:16.960 --> 00:45:19.880 explain this to me as if I was a 6-year-old? LEE: Yeah. [LAUGHS]  00:45:19.880 --> 00:45:24.640 CHEN: Breaking down complex topics. Yeah.  So I think it's something that we see in   00:45:24.640 --> 00:45:29.200 the pre-clinical space, in lecture,  but also even in the clinical space,   00:45:30.320 --> 00:45:34.720 there's a lot of teaching, as well. Sometimes if my preceptor is busy with patients,   00:45:34.720 --> 00:45:39.840 but I had maybe a question, I would maybe converse  with ChatGPT, like, “Hey, what are your thoughts   00:45:39.840 --> 00:45:44.940 about this?” Or, like, a common one is, like,  medical doctors love to use abbreviations, …  00:45:44.940 --> 00:45:46.000 LEE: Yes. CHEN: … and these abbreviations   00:45:46.000 --> 00:45:50.160 are sometimes only very niche and unique  to their specialty, right. [LAUGHTER]  00:45:50.160 --> 00:45:54.720 And I was reading this note from a  urogynecologist. [In] the entire first sentence,   00:45:54.720 --> 00:45:59.360 I think there were, like, 10 abbreviations.  Obviously, I compile lists and ask ChatGPT,   00:45:59.360 --> 00:46:04.320 like, “Hey, in the context of urogynecology, can  you define what these could possibly mean,” right?   00:46:04.320 --> 00:46:09.280 Instead of hopelessly searching in a Google  or maybe, embarrassing, asking the preceptor.   00:46:09.280 --> 00:46:16.640 So in these instances, it's played a huge role. LEE: Yeah. And when you're doing things like that,   00:46:16.640 --> 00:46:24.800 it can make mistakes. And so what  are your views of the reliability of   00:46:24.800 --> 00:46:31.600 generative AI, at least in the form of ChatGPT? CHEN: Yeah, I think into the context of medicine,   00:46:31.600 --> 00:46:37.440 right, we fear a lot about the hallucinations  that these models might have. And it's something   00:46:37.440 --> 00:46:42.720 I'm always checking for. When I talk with  peers about this, we find it most helpful   00:46:42.720 --> 00:46:50.240 when the model gives us a source linking it  back. I think the gold standard nowadays in   00:46:50.240 --> 00:46:54.880 medicine is using something called UpToDate  that's written by clinicians, for clinicians.  00:46:56.080 --> 00:47:01.680 But sometimes searching on UpToDate can be a lot  of time as well because it's a lot of information   00:47:01.680 --> 00:47:06.880 to, like, sort through. But nowadays a lot of  us are using something called OpenEvidence,   00:47:06.880 --> 00:47:13.760 which is also an LLM. But they always cite their  citations with, like, published literature, right.  00:47:13.760 --> 00:47:20.560 So I think being able to be conscious of the  downfalls of these models and also being able   00:47:20.560 --> 00:47:24.560 to have the critical understanding of,  like, analyzing the actual literature.   00:47:24.560 --> 00:47:28.800 I think double checking is just something  that we've been also getting really good at.  00:47:28.800 --> 00:47:39.280 LEE: How would you assess student attitudes—med  student attitudes—about AI? Is it … the way you're   00:47:39.280 --> 00:47:45.360 coming across is it's just a natural part  of life. But do people have firm opinions,   00:47:45.360 --> 00:47:51.680 you know, pro or con, when it comes  to AI, and especially AI in medicine?  00:47:51.680 --> 00:47:58.320 CHEN: I think it's pretty split right  now. I think there's the half, kind of,   00:47:58.320 --> 00:48:04.080 like us, where we're very optimistic—cautiously  optimistic about, you know, the potential of this,   00:48:04.080 --> 00:48:10.480 right. It's able to, you know, give us that extra  information, of being that extra tutor, right.   00:48:11.120 --> 00:48:15.280 It's also able to give us  information very quickly, as well.  00:48:15.280 --> 00:48:20.160 But I think the other flip side of what a  lot of students hesitate to, which I agree,   00:48:20.160 --> 00:48:29.040 is this loss of the ability to critically think.  Something that you can easily do is, you know,   00:48:29.040 --> 00:48:34.960 give these models, like, relevant information  about the patient history and be like, “Give me   00:48:34.960 --> 00:48:37.760 a 10-list differential,” right LEE: Yes.  00:48:37.760 --> 00:48:41.200 CHEN: And I think it's very easy as a  student to, you know, [say], “This is   00:48:41.200 --> 00:48:45.280 difficult. Let me just use what the model  says, and we'll go with that,” right.  00:48:45.280 --> 00:48:51.840 So I think being able to separate that,  you know, medical school is a time where,   00:48:51.840 --> 00:48:56.000 you know, you're learning to become a good  doctor. And part of that requires the ability   00:48:56.000 --> 00:49:03.120 to be observant and critically think. Having  these models simultaneously might hinder the   00:49:03.120 --> 00:49:04.210 ability to do that. LEE: Yeah.  00:49:04.210 --> 00:49:09.840 CHEN: So I think, you know, the next step is,  like, these models can be great—a great tool,   00:49:09.840 --> 00:49:13.520 absolutely wonderful. But how do you  make sure that it's not hindering   00:49:13.520 --> 00:49:18.960 these abilities to critically

think? LEE: Right. And so when you're doing your   00:49:18.960 --> 00:49:23.760 LIC [longitudinal integrated clerkship]  work, these longitudinal experiences,   00:49:23.760 --> 00:49:30.360 and you're in clinic, are you pulling the phone  out of your pocket and consulting with AI?  00:49:30.360 --> 00:49:35.600 CHEN: Definitely. And I think my own policy  for this, to kind of counter this, is that the   00:49:35.600 --> 00:49:39.520 night before when I'm looking over the patient  list, the clinic [schedule] of who's coming,   00:49:39.520 --> 00:49:44.960 I'm always giving it my best effort first. Like, OK, the chief complaint is maybe just a   00:49:44.960 --> 00:49:50.080 runny nose for a kid in a pediatric clinic. What  could this possibly be? Right? At this point,   00:49:50.080 --> 00:49:52.400 we've seen a lot. Like, OK, it could be URI  [upper respiratory infection], it could be   00:49:52.400 --> 00:49:55.920 viral, it could be bacterial, you know,  and then I go through the—you know, I try   00:49:55.920 --> 00:50:00.160 to do my due diligence of, like, going through  the history and everything like that, right.  00:50:00.160 --> 00:50:04.400 But sometimes if it's a more complex case,  something maybe a presentation I've never   00:50:04.400 --> 00:50:09.760 seen before, I'll still kind of do my best  coming up with maybe a differential that might   00:50:09.760 --> 00:50:15.360 not be amazing. But then I'll ask, you know,  ChatGPT like, OK, in addition to these ideas,   00:50:15.360 --> 00:50:16.314 what do you think? LEE: Yeah.  00:50:16.314 --> 00:50:21.480 CHEN: Am I missing something? You know, and  usually, it gives a pretty good response.  00:50:21.480 --> 00:50:26.160 LEE: You know, that particular idea  is something that I think Carey, Zak,   00:50:26.160 --> 00:50:31.120 and I thought would be happening a lot more  today than we're observing. And it's the idea of   00:50:31.120 --> 00:50:37.920 a second set of eyes on your work. And somehow, at  least our observation is that that isn't happening   00:50:37.920 --> 00:50:43.280 quite as much by today as we thought it might. And it just seems like one of the really safest   00:50:43.280 --> 00:50:48.640 and most effective use cases. When  you go and you're looking at yourself   00:50:48.640 --> 00:50:54.080 and other fellow medical students, other  second-year students, what do you see when   00:50:54.080 --> 00:51:01.760 it comes to the “second set of eyes” idea? CHEN: I think, like, a lot of students are   00:51:01.760 --> 00:51:09.840 definitely consulting ChatGPT in that regard  because, you know, even in the very beginning,   00:51:09.840 --> 00:51:16.080 we're taught to be, like, never miss these red  flags, right. So these red flags are always on our   00:51:16.080 --> 00:51:22.320 differential, but sometimes, it can be difficult  to figure out where to place them on that, right.  00:51:22.320 --> 00:51:27.840 So I think in addition to, you know,  coming up with these differentials,   00:51:27.840 --> 00:51:33.120 something I've been finding a lot of value [in]  is just chatting with these tools to get their   00:51:33.120 --> 00:51:38.960 rationale behind their thinking, you know. Something I find really helpful—I think   00:51:38.960 --> 00:51:44.080 this is also a part of the, kind of, art of  medicine—is figuring out what to order, right,   00:51:44.080 --> 00:51:45.354 what labs to order. LEE: Right.  00:51:45.354 --> 00:51:48.400 CHEN: Obviously, you have your order  sets that automate some of the things,   00:51:48.400 --> 00:51:51.520 like in the ED [emergency department], or, like,  there are some gold standard imaging things   00:51:51.520 --> 00:51:55.600 you should do for certain presentations. But then you chat to, like, 10 different   00:51:55.600 --> 00:51:59.000 physicians on maybe the next steps after  that, and they give you 10 different answers.  00:51:59.000 --> 00:52:00.800 LEE: Yes. CHEN: But there's never ... I   00:52:00.800 --> 00:52:05.200 never understand exactly why. It's always like,  I've just been doing this for all my training,   00:52:05.200 --> 00:52:09.280 or that's how I was taught. So asking ChatGPT, like,   00:52:09.280 --> 00:52:12.880 “Why would you do this next?” Or, like, “Is  this a good idea?” And seeing the pros and   00:52:12.880 --> 00:52:20.720 cons has also been really helpful in my learning. LEE: Yeah, wow, that's super interesting. So now,   00:52:21.440 --> 00:52:27.600 you know, I'd like to get into the education  you're receiving. And, you know, I think it's fair   00:52:27.600 --> 00:52:35.680 to say Kaiser Permanente is very progressive  in really trying to be very cutting-edge   00:52:35.680 --> 00:52:42.800 in how the whole curriculum is set up. And for the listeners who don't know this,   00:52:42.800 --> 00:52:48.960 I'm actually on the board of directors of the  school and have been since the founding of the   00:52:48.960 --> 00:52:54.640 school. And I think one of the reasons  why I was invited to be on the board is   00:52:55.520 --> 00:53:02.640 the school really wanted to think ahead and  be cutting edge when it comes to technology.  00:53:03.600 --> 00:53:11.600 So from where I've sat, I've never been completely  satisfied with the amount of tech that has made it   00:53:11.600 --> 00:53:19.200 into the curriculum. But at the same time, I've  also made myself feel better about that just   00:53:19.200 --> 00:53:24.720 understanding that it's sort of unstoppable,  that students are so tech-forward already.  00:53:25.520 --> 00:53:32.960 But I wanted to delve into a little bit here  into what your honest opinions are and your   00:53:32.960 --> 00:53:38.960 fellow students' opinions are about whether  you feel like you're getting adequate training   00:53:38.960 --> 00:53:44.640 and background formally as part of your medical  education when it comes to things like artificial   00:53:44.640

--> 00:53:49.280 intelligence or other technologies. What do you think? Are you   00:53:49.280 --> 00:53:55.360 … would you wish the curriculum would change? CHEN: Yeah, I think that's a great question.  00:53:55.360 --> 00:54:01.280 I think from a tech perspective, the school  is very good about implementing, you know,   00:54:01.280 --> 00:54:08.000 opportunities for us to learn. Like, for  example, learning how to use Epic, right,   00:54:08.000 --> 00:54:11.600 or at Kaiser Permanente, what we call  HealthConnect, right. These electronic   00:54:11.600 --> 00:54:15.920 health records. That, my understanding is,  a lot of schools maybe don't teach that.  00:54:15.920 --> 00:54:21.440 That's something where we get training sessions  maybe once or twice a year, like, “Hey, here's how   00:54:21.440 --> 00:54:26.880 to make a shortcut in the environment,” right. So I think from that perspective, the school   00:54:26.880 --> 00:54:32.320 is really proactive in providing those  opportunities, and they make it very easy   00:54:32.320 --> 00:54:37.440 to find resources for that, too. I think it ... LEE: Yeah, I think you're pretty much guaranteed   00:54:37.440 --> 00:54:40.240 to be an Epic black belt by  the time you [LAUGHS] finish   00:54:40.240 --> 00:54:42.800 your degree. CHEN: Yes, yes.  00:54:42.800 --> 00:54:48.480 But then I think in terms of the aspects of  artificial intelligence, I think the school's   00:54:48.480 --> 00:54:53.920 taken a more cautiously optimistic viewpoint.  They're just kind of looking around right now.  00:54:54.640 --> 00:54:59.200 Formally in the curriculum, there hasn't  been anything around this topic. I believe   00:54:59.200 --> 00:55:04.160 the fourth-year students last year got  a student-led lecture around this topic.  00:55:04.800 --> 00:55:11.760 But talking to other peers at other institutions,  it looks like it's something that's very slowly   00:55:11.760 --> 00:55:18.240 being built into the curriculum, and it seems like  a lot of it is actually student-led, you know.  00:55:18.960 --> 00:55:22.080 You know, my friend at Feinberg  [School of Medicine] was like we   00:55:22.080 --> 00:55:27.440 just got a session before clerkship about  best practices on how to use these tools.  00:55:28.080 --> 00:55:32.640 I have another friend at Pitt talking  about how they’re leading efforts of   00:55:32.640 --> 00:55:40.080 maybe incorporating some sort of LLM into  their in-house curriculum where students can,   00:55:40.080 --> 00:55:44.880 instead of clicking around the website trying to  find the exact slide, they can just ask this tool,   00:55:44.880 --> 00:55:48.720 like, “OK. We had class this day. They talked  about this ... but can you provide more   00:55:48.720 --> 00:55:52.560 information?” and it can pull from that. So I think a lot of this, a lot of it is   00:55:52.560 --> 00:55:59.200 student-driven. Which I think is really exciting  because it begs the question, I think, you know,   00:55:59.200 --> 00:56:04.080 current physicians may not be very well  equipped with these tools as well, right?  00:56:04.080 --> 00:56:07.920 So maybe they don't have a good idea  of what exactly is the next steps or   00:56:07.920 --> 00:56:11.840 what does the curriculum look like. So  I think the future in terms of this AI   00:56:11.840 --> 00:56:16.160 curriculum is really student-led, as well. LEE: Yeah, yeah, it's really interesting.  00:56:16.160 --> 00:56:21.440 I think one of the reasons I  think also that that happens is   00:56:21.440 --> 00:56:28.400 [that] it's not just necessarily the curriculum  that lags but the accreditation standards.   00:56:28.400 --> 00:56:33.200 You know, accreditation is really important  for medical schools because you want to make   00:56:33.200 --> 00:56:40.480 sure that anyone who holds an MD, you know,  is a bona fide doctor, and so accreditation   00:56:40.480 --> 00:56:47.200 standards are pretty strictly monitored in  most countries, including the United States.  00:56:47.200 --> 00:56:56.480 And I think accreditation standards are also—my  observation—are slow to understand how to adopt or   00:56:56.480 --> 00:57:03.440 integrate AI. And it's not meant as a criticism.  It's a big unknown. No one knows exactly   00:57:03.440 --> 00:57:10.640 what to do and how to do. And so it's really  interesting to see that, as far as I can tell,   00:57:10.640 --> 00:57:17.200 I've observed the same thing that you just have  seen, that most of the innovation in this area   00:57:17.200 --> 00:57:22.800 about how AI should be integrated into medical  education is coming from the students themselves.  00:57:22.800 --> 00:57:31.520 It seems, I think, I'd like to think  it's a healthy development. [LAUGHS]  00:57:31.520 --> 00:57:36.720 CHEN: Something tells me maybe the students  are a bit better at using these tools, as well.  00:57:36.720 --> 00:57:41.595 You know, I talk to my preceptors because KP  [Kaiser Permanente] also has their own version …  00:57:41.595 --> 00:57:44.221 LEE: Preceptor, maybe we  should explain what that is.  00:57:44.221 --> 00:57:49.280 CHEN: Yeah, sorry. So a preceptor is an attending  physician, fully licensed, finished residency,   00:57:50.000 --> 00:57:54.800 and they are essentially your kind of  teacher in the clinical environment.  00:57:56.000 --> 00:58:01.840 So KP has their own version of some ambient  documentation device, as well. And something   00:58:01.840 --> 00:58:06.720 I always like to ask, you know, like, “Hey,  what are your thoughts on these tools,” right?  00:58:06.720 --> 00:58:13.840 And it's always so polarizing, as well, even  among the same specialty. Like, if you ask   00:58:13.840 --> 00:58:18.880 psychiatrists, which I think is a great use case  of these tools, right. My preceptor hates it.   00:58:18.880 --> 00:58:22.960 Another preceptor next door loves it. [LAUGHTER] So I think a lot of it’s, like, it’s

still, like,   00:58:22.960 --> 00:58:27.360 a lot of unknowns, like you were mentioning. LEE: Right. Well, in fact, I'm glad you brought   00:58:27.360 --> 00:58:32.160 that up because one thing that we've  been hearing from previous guests a lot   00:58:32.160 --> 00:58:40.320 when it comes to AI in clinic is about ambient  listening by AI, for example, to help set up a   00:58:40.320 --> 00:58:46.160 clinical note or even write a clinical note. And another big use case that we heard a lot   00:58:46.160 --> 00:58:54.320 about that seems to be pretty popular is the use  of generative AI to respond to patient messages.  00:58:54.320 --> 00:59:03.360 So let's start with the clinical note thing. First  off, do you have opinions about that technology?  00:59:03.360 --> 00:59:07.280 CHEN: I think it's definitely good. I think especially where, you know, if you're   00:59:07.280 --> 00:59:11.200 in the family medicine environment or pediatric  environment where you're spending so much time   00:59:11.200 --> 00:59:19.040 with patients, a note like that is great, right. I think coming from a strictly medical student   00:59:19.040 --> 00:59:26.720 standpoint, I think it's—honestly, it'd be great  to have—but I think there's a lot of learning when   00:59:26.720 --> 00:59:31.920 you write the note, you know. There's a lot of,  you know, all of my preceptors talk about, like,   00:59:31.920 --> 00:59:37.440 when I read your note, you should present it in  a way where I can see your thoughts and then once   00:59:37.440 --> 00:59:41.520 I get to the assessment and plan, it's kind of  funneling down towards a single diagnosis or a   00:59:41.520 --> 00:59:48.560 handful of diagnoses. And that's, I think, a skill  that requires you to practice over time, right.  00:59:48.560 --> 00:59:53.520 So a part of me thinks, like, if I had this  tool where [it] can just automatically give me   00:59:53.520 --> 01:00:00.560 a note as a first year, then it takes away  from that learning experience, you know.  01:00:01.200 --> 01:00:06.560 Even during our first year throughout  school, we frequently get feedback from   01:00:07.200 --> 01:00:12.000 professors and doctors about these notes. And  it's a lot of feedback. [LAUGHTER] It's like,   01:00:12.000 --> 01:00:15.040 “I don't think you should have written that,”  “That should be in this section ” … you know,   01:00:15.040 --> 01:00:18.640 like a medical note or a SOAP note [Subjective,  Objective, Assessment, and Plan], where, you   01:00:18.640 --> 01:00:24.080 know, the subjective is, like, what the patient  tells you. Objective is what the physical findings   01:00:24.080 --> 01:00:28.960 are, and then your assessment of what's happening,  and then your plan. Like, it's very particular,   01:00:28.960 --> 01:00:34.000 and then I think medicine is so structured in a  way, that's kind of, like, how everyone does it,   01:00:34.000 --> 01:00:38.320 right. So kind of going back to the question,  I think it's a great tool, but I don't think   01:00:38.320 --> 01:00:41.760 it's appropriate for a medical student. LEE: Yeah, it's so interesting to hear   01:00:41.760 --> 01:00:50.000 you say that. I was … one of our previous guests  is the head of R&D at Epic, Seth Hain. He said,   01:00:50.000 --> 01:00:56.880 “You know, Peter, doctors do a lot of  their thinking when they write the note.”  01:00:56.880 --> 01:01:03.680 And, of course, Epic is providing ambient, you  know, clinical notetaking automation. But he was   01:01:03.680 --> 01:01:08.480 urging caution because, you know, you're  saying, well, this is where you're learning   01:01:08.480 --> 01:01:14.400 a lot. But actually, it's also a point where, as  a doctor, you're thinking about the patient. And   01:01:16.080 --> 01:01:22.080 we do probably have to be careful  with how we automate parts of that.  01:01:22.080 --> 01:01:27.746 All right. So you're gearing up for Step 1  of the USMLE [United States Medical Licensing   01:01:27.746 --> 01:01:34.080 Examination]. That'll be a big multiple-choice  exam. Then Step 2 is similar: very, very focused   01:01:34.080 --> 01:01:39.760 on advanced clinical knowledge. And then Step  3, you know, is a little more interactive.  01:01:41.920 --> 01:01:50.080 And so one question that people have had  about AI is, you know, how do we regulate   01:01:50.080 --> 01:02:01.440 the use of AI in medicine? And one of the famous  papers that came out of both academia and industry   01:02:01.440 --> 01:02:07.920 was the concept that you might be able to treat  AI like a person and have it go through the   01:02:07.920 --> 01:02:15.680 same licensing. And this is something that  Carey, Zak, and I contemplated in our book.  01:02:16.880 --> 01:02:21.760 In the end, at the time we wrote the  book, I personally rejected the idea,   01:02:21.760 --> 01:02:29.680 but I think it's still alive. And so I’ve  wondered if you have any ... you know, first off,   01:02:29.680 --> 01:02:36.960 are you opinionated at all about, what should  the requirements be for the allowable use of AI   01:02:36.960 --> 01:02:41.520 in the kind of work that you're going to be doing? CHEN: Yeah, I think it's a tough question because,   01:02:41.520 --> 01:02:47.200 like, where do you draw that line, right? If you  apply the human standards of it's passing exams,   01:02:47.200 --> 01:02:50.800 then yes, in theory, it could be  maybe a medical doctor, as well,   01:02:50.800 --> 01:02:55.120 right? It's more empathetic than medical  doctors, right? So where do you draw that line?  01:02:55.120 --> 01:03:00.000 I think, you know, part of me thinks  it's maybe it is that human aspect   01:03:00.000 --> 01:03:05.920 that patients like to connect with,  right. And maybe this really is just,   01:03:05.920 --> 01:03:12.400 like, these tools are just aids in helping, you  know, maybe load off some cognitive load, right.  01:03:12.400 --> 01:03:18.400 But

I think the other part of me, I'm thinking  about this is the next generation who are growing   01:03:18.400 --> 01:03:23.120 up with this technology, right. They're  interacting with applications all day.   01:03:23.120 --> 01:03:28.160 Maybe they're on their iPads. They're talking  to chatbots. They're using ChatGPT. This is,   01:03:28.160 --> 01:03:34.080 kind of, the environment they grew up with.  Does that mean they also have increased,   01:03:34.080 --> 01:03:40.080 like, trust in these tools that maybe our  generation or the generations above us don't   01:03:40.080 --> 01:03:44.480 have that value that human connection?  Would they value human connection less?  01:03:44.480 --> 01:03:48.240 You know, I think those are some troubling  thoughts that, you know, yes, at end of the day,   01:03:48.240 --> 01:03:53.040 maybe I'm not as smart as these tools, but I  can still provide that human comfort. But if,   01:03:53.040 --> 01:03:57.680 at the end of the day, the future generation  doesn't really care about that or they perfectly   01:03:57.680 --> 01:04:03.200 trust these tools because that's all they've  kind of known, then where do human doctors stand?  01:04:03.840 --> 01:04:10.240 I think part of that is, there would be certain  specialties where maybe the human connection is   01:04:10.240 --> 01:04:14.400 more important. The longitudinal aspect of  building that trust, I think is important.   01:04:14.400 --> 01:04:20.160 Family medicine is a great example. I think  hematology oncology with cancer treatment.  01:04:20.160 --> 01:04:25.440 Obviously, I think anyone's not going  to be thrilled to hear cancer diagnosis,   01:04:25.440 --> 01:04:29.920 but something tells me that seeing that on a  screen versus maybe a physician prompting you   01:04:29.920 --> 01:04:37.280 and telling you about that tells me that maybe  in those aspects, you know, the human nature, the   01:04:37.280 --> 01:04:46.080 human touch plays an important role there, too. LEE: Yeah, you know, I think it strikes me that   01:04:46.080 --> 01:04:53.200 it's going to be your generation that really  is going to set the pattern probably for the   01:04:53.200 --> 01:05:02.560 next 50 years about how this goes. And it's  just so interesting because I think a lot   01:05:02.560 --> 01:05:07.120 will depend on your reactions to things. So, for example, you know, one thing that   01:05:07.120 --> 01:05:12.160 is already starting to happen are  patients who are coming in armed,   01:05:12.160 --> 01:05:17.600 you know, with a differential [LAUGHS], you  know, that they've developed themselves with   01:05:17.600 --> 01:05:23.600 the help of ChatGPT. So let me ... you must  have thought about these things. So, in fact,   01:05:23.600 --> 01:05:29.280 has it happened in your clinical work already? CHEN: Yeah, I've seen people come into the   01:05:29.280 --> 01:05:35.760 ED during my ED shift, like emergency  department, and they’ll be like, “Oh,   01:05:35.760 --> 01:05:40.800 I have neck pain and here are all the things  that, you know, Chat told me, ChatGPT told   01:05:40.800 --> 01:05:45.440 me. What do you think ... do I need? I  want this lab ordered, that lab ordered.”  01:05:45.440 --> 01:05:47.040 LEE: Right. CHEN: And I think my   01:05:47.040 --> 01:05:52.720 initial reaction is, “Great. Maybe we should  do that.” But I think the other reaction is   01:05:52.720 --> 01:06:00.960 understanding that not everyone has the clinical  background of understanding what's most important,   01:06:00.960 --> 01:06:09.200 what do we need to absolutely rule out, right? So, I think in some regards, I would think that   01:06:09.200 --> 01:06:11.990 maybe ChatGPT errs on the side of caution, ... LEE: Yes.  01:06:11.990 --> 01:06:16.000 CHEN: ... giving maybe patients more  extreme examples of what this could be   01:06:16.000 --> 01:06:22.456 just to make sure that it's, in a way, is  not missing any red flags as well, right.  01:06:22.456 --> 01:06:26.240 LEE: Right. Yeah. CHEN: But I think a lot of this is … what   01:06:26.240 --> 01:06:30.720 we've been learning is it's all about shared  decision making with the patient, right. Being   01:06:30.720 --> 01:06:36.880 able to acknowledge like, “Yeah, [in] that list,  most of the stuff is very plausible, but maybe   01:06:36.880 --> 01:06:43.440 you didn't think about this one symptom you have.” So I think part of it, maybe it's a sidebar here,   01:06:43.440 --> 01:06:48.320 is the idea of prompting, right. You know,  they've always talked about all these, you know,   01:06:48.320 --> 01:06:54.080 prompt engineers, you know, how well can you,  like, give it context to answer your question?  01:06:54.080 --> 01:06:55.120 LEE: Yeah. CHEN: So I think being   01:06:55.120 --> 01:07:00.480 able to give these models the correct information  and the relevant information and keyword relevant,   01:07:00.480 --> 01:07:05.840 because relevant is, I guess, where your clinical  expertise comes in. Like, what do you give the   01:07:05.840 --> 01:07:10.960 model, what do you not give? So I think that  difference between a medical provider versus   01:07:10.960 --> 01:07:16.240 maybe your patients is ultimately the difference. LEE: Let me press on that a little bit more   01:07:16.240 --> 01:07:23.440 because you brought up the issue of trust,  and trust is so essential for patients to   01:07:23.440 --> 01:07:31.200 feel good about their medical care. And I can imagine you're a medical   01:07:31.200 --> 01:07:36.560 student seeing a patient for the first time.  So you don't have a trust relationship with   01:07:36.560 --> 01:07:43.200 that patient. And the patient comes in  maybe trusting ChatGPT more than you.  01:07:43.200 --> 01:07:47.920 CHEN: Very valid. No. I mean, I  get that a lot, surprisingly, you   01:07:47.920 --> 01:07:51.280 know. [LAUGHTER] Sometimes [they’re] like, “Oh, I  don't want to see the medical

student,” because we   01:07:51.280 --> 01:07:55.520 always give the patient an option, right. Like,  it's their time, whether it's a clinic visit.  01:07:56.240 --> 01:08:01.360 But yeah, those patients, I think it's  perfectly reasonable. If I heard a   01:08:01.360 --> 01:08:05.200 second-year medical student was going to be  part of my care team, taking that history,   01:08:05.200 --> 01:08:09.840 I'd be maybe a little bit concerned, too.  Like, are they asking all the right questions?   01:08:09.840 --> 01:08:15.120 Are they relaying that information back  to their attending physician correctly?  01:08:15.840 --> 01:08:22.160 So I think a lot of it is, at least from a medical  student perspective, is framing it so the patient   01:08:22.160 --> 01:08:29.840 understands that this is a learning opportunity  for the students. And something I do a lot is tell   01:08:29.840 --> 01:08:34.160 them like, “Hey, like, you know, at the end  of the day, there is someone double-checking   01:08:34.160 --> 01:08:35.360 all my work.” LEE: Yeah.  01:08:35.360 --> 01:08:39.600 CHEN: But for those that come in with  a list, I sometimes sit down with them,   01:08:39.600 --> 01:08:44.720 and we'll have a discussion, honestly. I’ll be like, “I don't think you have   01:08:44.720 --> 01:08:50.320 meningitis because you're not having a fever.  Some of the physical exam maneuvers we did were   01:08:50.320 --> 01:08:54.960 also negative. So I don't think you have  anything to worry about that,” you know.  01:08:54.960 --> 01:09:00.160 So I think it's having that very candid  conversation with the patient that helps build   01:09:00.160 --> 01:09:04.160 that initial trust. Telling them like, “Hey … ” LEE: It's impressive to hear how even keeled you   01:09:04.160 --> 01:09:09.200 are about this. You know, I think, of course, and  you're being very humble saying, well, you know,   01:09:09.200 --> 01:09:13.440 as a second-year medical student, of course,  someone might not, you know, have complete   01:09:13.440 --> 01:09:21.200 trust. But I think that we will be entering  into a world where no doctor is going to be,   01:09:21.200 --> 01:09:28.560 no matter how experienced or how skilled,  is going to be immune from this issue.  01:09:28.560 --> 01:09:32.960 So we're starting to run toward the end  of our time together. And I like to end   01:09:32.960 --> 01:09:41.440 with one or two more provocative questions. And so let me start with this one. Undoubtedly,   01:09:41.440 --> 01:09:46.160 I mean, you're close enough to tech  and digital stuff, digital health,   01:09:46.160 --> 01:09:52.160 that you're undoubtedly familiar  with famous predictions, you know,   01:09:52.160 --> 01:10:01.760 by Turing and Nobel laureates that someday certain  medical specialties, most notably radiology, would   01:10:01.760 --> 01:10:10.560 be completely supplanted by machines. And more  recently, there have been predictions by others,   01:10:10.560 -->

01:10:16.720 like, you know, Elon Musk, that maybe even some  types of surgery would be replaced by machines.  01:10:16.720 --> 01:10:22.560 What do you think? Do you have an opinion? CHEN: I think replace is a strong term,   01:10:22.560 --> 01:10:29.680 right. To say that doctors are completely  obsolete, I think, is unlikely.  01:10:29.680 --> 01:10:35.680 If anything, I think there might be a shift  maybe in what it means to be a doctor, right.   01:10:36.240 --> 01:10:41.600 Undoubtedly, maybe the demands of radiologists  are going to go down because maybe more of the   01:10:41.600 --> 01:10:45.360 simple things can truly be automated,  right. And you just have a supervising   01:10:45.360 --> 01:10:51.920 radiologist whose output is maybe 10 times  as maybe 10 single radiologists, right.  01:10:51.920 --> 01:10:56.560 So I definitely see a future where the  demand of certain specialties might go down.  01:10:56.560 --> 01:11:00.400 And I think when I talk about a shift  of what it means to be a physician,   01:11:00.400 --> 01:11:05.680 maybe it's not so much diagnostic anymore,  right, if these models get so good at, like,   01:11:05.680 --> 01:11:11.440 just taking in large amounts of information,  but maybe it pivots to being really good at   01:11:11.440 --> 01:11:16.880 understanding the limitations of these models and  knowing when to intervene is what it means to be   01:11:16.880 --> 01:11:23.920 the kind of the next generation of physicians. I think in terms of surgery, yeah, I think it's   01:11:23.920 --> 01:11:29.920 a concern, but maybe not in the next 50 years.  Like those Da Vinci robots are great. I think out   01:11:29.920 --> 01:11:36.880 of Mayo Clinic, they were demoing some videos of  these robots leveraging computer vision to, like,   01:11:36.880 --> 01:11:40.880 close portholes, like laparoscopic scars. And  that's something I do in the OR [operating room],   01:11:40.880 --> 01:11:45.600 right. And we're at the same level at this  point. [LAUGHTER] So at that point, maybe.  01:11:45.600 --> 01:11:49.360 But I think robotics still has to  address the understanding of like,   01:11:49.360 --> 01:11:54.400 what if something goes wrong, right? Who's  responsible? And I don't see a future where   01:11:54.400 --> 01:12:00.480 a robot is able to react to these, you know,  dangerous situations when maybe something   01:12:00.480 --> 01:12:07.520 goes wrong. You still have to have a surgeon on  board to, kind of, take over. So in that regard,   01:12:07.520 --> 01:12:16.640 that's kind of where I see maybe the future going. LEE: So last question. You know, when you are   01:12:16.640 --> 01:12:26.640 thinking about the division of time, one of the  themes that we've seen in the previous guests is   01:12:26.640 --> 01:12:35.760 more and more doctors are doing more technology  work, like writing code and so on. And more   01:12:35.760 --> 01:12:46.080 and more technologists are thinking deeply and  getting educated in clinical and preclinical work.  01:12:46.080 --> 01:12:54.320 So for you, let's look ahead 10 years. What  do you see your division of labor to be? Or,   01:12:54.320 --> 01:12:59.440 you know, how would you … what would you tell  your mom then about how you spend a typical day?  01:12:59.440 --> 01:13:05.520 CHEN: Yeah, I mean, I think for me, technology  is something I definitely want to be involved   01:13:05.520 --> 01:13:11.280 in in my line of work, whether it's, you  know, AI work, whether it's improving   01:13:11.280 --> 01:13:17.600 quality of healthcare through technology. My perfect division would be maybe still   01:13:17.600 --> 01:13:22.800 being able to see patients but also balancing  some maybe more of these higher-level kind   01:13:22.800 --> 01:13:26.920 of larger projects. But I think having  that division would be something nice.  01:13:26.920 --> 01:13:34.480 LEE: Yeah, well, I think you would be great  just from the little bit I know about you. And,   01:13:34.480 --> 01:13:39.840 Daniel, it's been really great chatting with  you. I wish you the best of luck, you know,   01:13:39.840 --> 01:13:46.160 with your upcoming exams and getting past  this year two of your medical studies. And   01:13:47.760 --> 01:13:50.314 perhaps someday I'll be your patient. [TRANSITION MUSIC]  01:13:50.314 --> 01:13:55.320 CHEN: Thank you so much. LEE:   01:14:00.480 --> 01:14:05.600 You know, one of the lucky things about my job is  that I pretty regularly get to talk to students   01:14:05.600 --> 01:14:12.240 at all levels, spanning high school to graduate  school. And when I get to talk especially to med   01:14:12.240 --> 01:14:17.360 students, I'm always impressed with their  intelligence, just how serious they are,   01:14:17.360 --> 01:14:23.600 and their high energy levels. Daniel is  absolutely a perfect example of all that.  01:14:23.600 --> 01:14:29.520 Now, it comes across as trite to say that the  older generation is less adept at technology   01:14:29.520 --> 01:14:36.160 adoption than younger people. But actually,  there probably is a lot of truth to that. And   01:14:36.160 --> 01:14:40.880 in the conversation with Daniel, I think he  was actually being pretty diplomatic but also   01:14:40.880 --> 01:14:48.160 clear that he and his fellow med students don't  necessarily expect the professors in their med   01:14:48.160 --> 01:14:53.120 school to understand AI as well as they do. There's no doubt in my mind that medical   01:14:53.120 --> 01:14:57.680 education will have to evolve a lot to  help prepare doctors and nurses for an AI   01:14:57.680 --> 01:15:04.240 future. But where will this evolution come from? As I reflect on my conversations with Morgan and   01:15:04.240 --> 01:15:11.360 Daniel, I start to think that it's most likely  to come from the students themselves. And when   01:15:11.360 --> 01:15:17.520 you meet people like Morgan and Daniel, it's  impossible to not be incredibly optimistic about   01:15:17.520 --> 01:15:19.920 the next generation of clinicians. [THEME MUSIC]  01:15:19.920 --> 01:15:33.920 Another big thank-you to Morgan and Daniel for  taking time to share their experiences with us.   01:15:33.920 --> 01:15:39.440 And to our listeners, thank you for joining  us. We have just a couple of episodes left,   01:15:39.440 --> 01:15:44.480 one on AI’s impact on the operation of public  health departments and healthcare systems and   01:15:44.480 --> 01:15:48.960 another coauthor roundtable. We  hope you'll continue to tune in.  01:15:48.960 --> 01:16:00.155 Until next time. [MUSIC FADES]