AI in Medicine: Colorado Medical Society

John Lantos

Thank you so much for inviting me. This is a topic that I think people are both a little bit sick of hearing about, but also a little bit concerned about. I'm going to take you through some thoughts that I've had about it. I'm a practicing pediatrician and have spent my career writing about medical ethics issues and doctor-patient relationship stuff. But Sean invited me to this conference based on an article that I wrote and published in JAMA in March called The Lost Aura of the Physician in the Age of Artificial Intelligence. Can you hear? Thank you. Let's see. Let me get this on a little better. Is that better? Yeah? Good. The Lost Aura is a little bit of a riff on the work of a German philosopher, Walter Benjamin, who was writing in the early part of the 20th century, mostly in the 1920s and the 1930s, and has a famous essay called The Work of Art in the Age of Mechanical Reproduction.

John Lantos

And what Benjamin was concerned about was the effect that photography and film would have on the visual arts, particularly on painting. He was writing at a time when film was making dramatic breakthroughs in the sorts of things it could talk about, and these were being politicized in ways that concerned Walter Benjamin. Hitler was famously a master of creating visual spectacles, filming them and using them for propaganda purposes that created a persona, one might say an aura about his political program and abilities. At the same time, Hollywood was experimenting with new critiques of capitalism in the works of Charlie Chaplin and others who, contra the Hitler films, was trying to undermine the tyrants and use this against them. And so Walter Benjamin was looking at these and thinking about what this was doing to the way we think about art, and he thought that art was once embedded in ritual and religious traditions, and the meaning of a work of art was inseparable from its embeddedness in this tradition and unique history.

John Lantos

He talked about how much of art was hung in churches or other places of worship and had a specific meaning in the places where it was hung. I was thinking about this in relation to what's happening today with doctor-patient relationships and AI, where in a sense what AI is doing is creating a mechanical or technological reproduction of the work that doctors do. To get a portrait, you used to need a painter. To get healthcare, you needed a doctor. And the doctor's aura, I argued, came from many things, but from our long training, our extensive clinical experience, the fact that we provided a human presence in the room with the patient, the ways that we would exercise clinical judgment. And the question was, could the machines ever replace these aspects the way a photograph could replace a painstakingly created oil paint portrait? The doctor's aura, medical expertise seems to be inseparable from the particular person who exercises it, and the in-the-moment encounter between a wise clinician and a suffering patient.

John Lantos

And a mechanical copy could be better in some ways. I mean, it's pretty clear that AI can know more than we do, can remember more than we do, can retrieve information from existing databases better than we do. It does better at getting to the diagnosis in New England Journal CPCs. It is perceived by patients as more empathic. So it might be better in some ways, but there's this faith that the copy would always be lacking in something essential and important, something comparable to Benjamin's aura that he associated with original works of art. So to recast his question, here's the question for today's talk. What becomes of that aura when the encounter is not with a human but with an algorithm? Or what remains of the aura when diagnosis becomes available on everybody's phone? So a lot of people are writing about this. Here are three of the people who I think are some of the smartest people writing about it.

John Lantos

Robert Wachter is head of medicine at UCSF. Peter Lee, I think, is at Microsoft. But Isaac Kohane is at Harvard and is now editor-in-chief of NEJM AI. Eric Topol is head of the Scripps Institute. And one of the things I found looking at these books, both when I was writing the Lost Aura paper and in preparing for this talk today, that sort of surprised me was how optimistic all three of these people are. You don't see any of the sort of dread that was making headlines this week. People quitting Anthropic in protest over their lack of controls. Congress having hearings about regulations. The Center for AI Safety rating this as a societal scale risk, such as a pandemic or a nuclear war warning that it will kill us all. And instead, what most of the doctors are saying is, this is going to be a valuable tool. We have other tools. We're going to learn to use it. It will work for us. We won't work for it. This is contrary to what many philosophers are saying about AI.

John Lantos

This is one of the most famous early warnings about AI. Nick Bostrom is a philosopher at Harvard. And he wrote an essay where he imagined giving AI an instruction to maximize the production of paper clips and suggested that without appropriate controls, AI would turn everything in the world into paper clips. It's sort of a sorcerer's apprentice kind of parable. With the conclusion that catastrophe doesn't come from AI consciously acting in an evil manner. It comes from a badly specified objective. If we're not careful about what we ask it to do, it may literally follow our instructions in ways that we didn't anticipate. By contrast, the leading physician organizations are like the writers, Wachter and Kohane and Topol. The AMA House of Delegates uses the term augmented intelligence rather than artificial intelligence, suggesting that it will augment our intelligence, emphasizing that its design enhances human intelligence but not replace it. The American College of Physicians has a similar sort of complacent attitude that technology should complement but not supplant the logic and decision-making of physicians.

John Lantos

I find this complacency a little alarming. I read it as evidence of just how thoroughly we've already made peace with our loss of aura. We seem to be happy to have become so much like machines that we're willing to use their intelligence to augment ours. We've done so gradually and perniciously and in ways that I think are deeply morally compromising and that AI will amplify. Let me talk about a few of the reasons I think that. In most of these books, there's four big common fears about AI. There's a hundred other fears about AI, including the doomsday scenarios. But in relation to medical AI, people talk about hallucinations. It makes things up, both facts and feelings like empathy. People talk about the problem of black box reasoning. That is, we don't understand how AI comes to the conclusions that it comes to. And even the people who design it don't know how it comes to the conclusions that it comes to. That's dangerous and we can't live with a technology that we don't understand in that deeply fundamental and important way.

John Lantos

People have talked a lot about de-skilling. That reliance on AI will lead us to lose the skills that we've so carefully developed. And finally, people have talked about institutional capture and the fear that we don't really know what the agendas of the tech companies that are developing this are. And they may capture us in ways that are disturbing. My own view is that all of these things are things that doctors themselves in the pre-AI era were doing more and more. And that AI is more of a mirror and an amplifier than it is a fundamentally disruptive transformation. Let me give a few examples. Humans make things up all the time. We call it theorizing. And in medicine, we come up with theories about what the best way to treat patients based on biologic plausibility and maybe some animal data and stuff like that. And much of the time, our theories are wrong.

John Lantos

But we implement them in patient care. In fact, there's a great book by Prasad and Sifu about what they call medical reversal, times when doctors confidently prescribed a certain treatment or intervention. And then, oops, somebody did a good study and found out that what we were recommending to our patients was, in fact, harmful. I'll just give one example from the book. The CAST trial for cardiac meds, who remembers when everybody after an MI would be put on antiarrhythmic medication? Why were they? Arrhythmias were thought to be life-threatening. We all assumed that this would be a good thing. And some people said, maybe it's not. A lot of doctors refused to participate in the randomized control trial because they said, we know it's beneficial. It would be unethical to randomize patients to placebo. But the study got done, and what they found was the patients on placebo had about a half or a third of the mortality rate of the patients who were getting the drug.

John Lantos

We could go through lists of other things, pulmonary artery catheters, PSA screening for everybody, postmenopausal hormone replacement, which has reversed and now is reversing again. But all these times we tell people with astounding confidence that we know the right thing to do. If AI did that, we'd say, you're making shit up, and we'd regulate it, and we'd ban it. But when humans do it, we somehow accept it. Empathy is also an interesting phenomenon here because, in a sense, for AI to express empathy is making up something deeply fundamental. I think whatever your concept of empathy is, an algorithm cannot feel it. And yet, in study after study, patients perceive AI's responses as more empathic than their human physician's responses. Sherry Turkle is a psychologist and sociologist at MIT who's been working with their AI group for 40 years and doing all sorts of studies, starting with kids and how kids would relate to early AI, if you want to call it that, not generative AI. But machines who expressed empathy, and the kids would bond with them and they would become friends.

John Lantos

It's now a common story in the news to talk about how people are developing deep and lasting emotional friendships with Claude or with ChatGPT. And Turkle came to the conclusion that we're deeply programmed to respond to somebody who seems to like us. And she says, we love what we allow ourselves to relate to. We're easy marks. We're cheap dates. If somebody says, I feel your pain. This must be really hard for you. You perceive that as loving and empathic, even if it's coming from a soulless machine. Did we get there first? Do humans do this as well? There's a great book called The Empathy Exams by Leslie Jamison. Jamison worked as a medical actor, giving exams to medical students where she was a simulated patient. And one of the things she was told to grade the students on was whether they expressed empathy. And in this grading, she was told it wouldn't be enough if they put their hand on your knee.

John Lantos

It wouldn't be enough if they have a kind look in their eyes. They have to express empathy. They have to say, this must be hard for you. They have to say, I can only imagine what you're going through. They were required to express empathy. And Jamison found this deeply troubling because it was immediately apparent to her that they were faking it. They were doing just what AI is doing and they'd get a passing grade. So the medical establishment was encouraging students to simulate empathy and not think about, she gets into, in the essay, and lots of other people have thought about this, what is real empathy? And how do we figure out ways to teach that? And in this essay, she said, it's in a book, but it's a chapter in the book. Empathy is not a feeling, it's a choice. Empathy isn't just listening. Empathy requires questioning, inquiry, and imagination. And empathy is not always a good thing. She says it can feel like a gift or it can feel like an invasion, particularly if through empathy you assume you understand what somebody is going through when they think you don't.

John Lantos

So whatever it would take to teach empathy, it's not what we've been doing with our medical students or with our AI. That's hallucinations. What about black box reasoning? Black box reasoning, as I say, is this idea that you put an input, you get an output, and you don't know how the machine got to the output, and people find that deeply troubling. But think about how you as physicians get to a diagnosis and whether you could explain what leaps you made in your mind from a rather sketchy history in a patient presenting with certain complaints to a diagnosis. Most people have trouble explaining their thought process. There's a great experiment they did at Harvard. It was like a Turing test for medical diagnosis. And it actually got written up in both NEJM AI by the doctor who participated in the Turing test, and in the New Yorker magazine by a doctor who writes a lot of the medical columns for them, Dhruv Khullar. What they did was gave AI and one of Harvard's best diagnosticians a case. And here was the basic outline. 41-year-old man, 10 days of fevers, body aches, swollen ankles, painful rash on his shins. He fainted twice. He had lung nodules and lymph nodes in his chest.

John Lantos

Dr. Restrepo, the human, had six weeks to prepare for this. And CaBot, the AI, had six minutes to prepare for this. They both got the correct diagnosis. And in this case, because it was a research study, the researchers had programmed CaBot to explain its thinking and its reasoning. And it turned out it got to the, and then they asked Restrepo, his reasoning. And they got to the diagnosis using completely different reasoning pathways. The things that Restrepo thought were the most important findings and that led him to the correct diagnosis were different from what the AI thought were the most important findings and led it to the correct diagnosis. So it raises the question, do we know how we think? And do we care how we think if we get to the right diagnosis? Can we know how we know what we know? Diagnosis and many other things in medicine are subtle forms of pattern recognition. We sift through mounds of data, make a sort of bet in our mind on which diagnosis seems likely, sift through emerging lab values and other things to see if they're consistent or inconsistent with the diagnosis.

John Lantos

Sometimes, even if they're consistent, we're not convinced. Sometimes if they're inconsistent, we ignore them for intuitive reasons. How exactly we do it and why some people are better is a mystery. Medical knowledge is knowing without knowing. Doctors know what they know, but not how they know it. And so we too are black boxes just like AI. And there's a whole literature trying to figure out how it is that doctors think, how the best doctors think, and how to teach. Students how to think like a doctor. Deskilling or deferring to machines and algorithms. Again, this has been going on for a long time, long before AI. Although AI is the new player in the room, particularly with ambient documentation. But medication decisions have been outsourced for a long time. Electrocardiograms, most of us rely on at least a preliminary reading by algorithmic machines. We used to be very good at examining abdomens. Now we do the point of care ultrasound. I know that because I watch The Pitt and it's the first thing they do when someone comes in. And AI stethoscopes now are starting to replace auscultation. And they do many of the things that ultrasounds do.

John Lantos

So one after another. I mean the neurologic exam is another favorite. You know, we used to do, neurologists used to do these detailed things to localize the lesion. Now, the neurologists I know won't even see the patient until they have the MRI. Sorry, neurologists. So there's two kinds of deskilling here to think about and worry about. One is the tool, and that's what a lot of those are. The AI stethoscope, the EKG computer, the point of care ultrasound. The other kind of deskilling that I think is maybe more interesting to think about is the protocol. So how often do we follow clinical guidelines? How often do we follow evidence-based medicine recommendations or algorithms in our clinical care? And let me give two examples to suggest that doing that may not be such a bad thing. Consider this pre-AI real-world study. It was done at the University of Utah in the 90s. It was a randomized control trial of ECMO versus conventional ventilator support for adults with ARDS. And they wanted to do this on patients who had very high expected mortality. So they figured out criteria for 90% mortality. Patients who met those criteria were eligible for the study.

John Lantos

And then they'd be randomized to ECMO or conventional mechanical ventilation. But when they were designing the study and started talking, they realized everybody had a different approach to mechanical ventilation. Some people went up on the rate. Some people went up on the oxygen. Some people went up on the PEEP. And there was no consensus. So before doing the study, they painstakingly went through case after case after case and got consensus. What should we do if the patient has X? So that the conventional arm of the therapy was a standardized intervention. And what they found when they did the randomized control trial was the ECMO patients had much better survival than the 90% predicted mortality. So did the patients in the control arm being treated by protocol rather than by individualized clinical judgment. They had a 60% mortality, the same as ECMO. If ECMO had been compared to historical controls, it would have looked much better. If you took away individualized clinical judgment and had the doctors follow a protocol, they did much better.

John Lantos

Form of de-skilling? Perhaps. It's a similar study in babies. The SUPPORT study randomized the babies to different target oxygen saturation levels. These are preemies in the NICU. They did a prospective randomized control trial. The doctors were masked to which baby was getting which oxygen target. And the interesting thing here, they found that lower oxygen had a higher mortality rate than higher oxygen. That was an interesting and controversial finding that I'm not going to discuss today. But what was also interesting and that nobody talked about as much is all the babies in the study did better than any reported series of babies at those gestational ages. So in the SUPPORT study, the babies in the low oxygen arm had a higher mortality rate. But overall, the mortality rate for every baby in the study was 18%. Every other study and registry data on similar babies had mortality rates over 20%. So in both studies, the protocol was better than the physician's individualized clinical judgment. AI is obviously going to be protocol driven.

John Lantos

This idea that we as experienced clinicians should always be able to override the protocol, I think is a good idea and is true. But I think having the guts, having the self-confidence to say my individualized clinical judgment is better than the protocol, when the data shows that in most cases protocols are better than individualized clinical judgment, is going to take a whole lot of confidence and may in fact lead to legal liability. Last thing, institutional capture. This is the one I am most, I guess, cynical about to worry that the tech bros are going to take over what we do, when in fact the medical profession and individual doctors have been institutionally captured in so many ways by so many different entities for the last, oh, 100 years. I think AI's effect is going to be more like rounding error than like a fundamental transformation. Think of professional regulation, think of government regulation, think of market influences. We are incentivized in every aspect of our clinical life to prioritize the things that generate revenue, even if those are not the things we find the most valuable.

John Lantos

Sandeep Jauhar wrote a great book about this called Doctored, the Disillusionment of an American Physician, in which he shows that doctors whose practices most align with the institution's incentives are the doctors who get promoted to positions of leadership, and that reinforces the idea that the institutional incentives and the professional incentives are the same thing. We've been captured in many ways. So, what are the implications of all this? For that, I'd like to go back to Walter Benjamin and think about what he predicted and what actually happened in the world of art in the 20th century and see whether an analogy to what's happening in medicine might help us think about this. What photography did for art? It didn't put art out of business, interestingly. Painters probably make a better living today than they did in 1920. What it did to art was change the focus of what the purpose of art was. It used to be much more about realism and creating illusions that you were seeing a landscape or a seascape or a person or a still life.

John Lantos

And what happened post-photography is art fragmented into all these exuberant new creative activities, impressionism, cubism, abstract expressionism, op art, conceptual art, urinals put on display, Campbell's soup cans as valuable works of art. Art suddenly had a new meaning that it never had had before. And it was something that people seemed to value and want or to disparage and hate. But it was not what art used to be. They were all previously unimaginable and very avant-garde. So I think the question to ask in this age of AI is can medicine have an avant-garde similar to art's avant-garde over the last century? And if so, what would that mean? What would it look like for medicine to enter an era where what it used to mean is not what it means anymore? And people start to experiment with new ways of thinking about what doctor-patient interactions should be. Well, here's one way to think about avant-garde medicine.

John Lantos

Consider two cutting-edge geneticists, leaders of their field, arguably. I never know how to pronounce this, but I'll say He Jiankui. He's a Chinese researcher. You may have read about him a few years ago when he decided to be trained in the United States, went back to China, and then edited the germline of two fetuses ostensibly to protect them from getting AIDS. He went to jail for that. He's now out of jail, and now he says he's going to continue trying to do his germline gene therapy experiments. In this case, he now wants to target the genes for Alzheimer's. It's going to be interesting to see what those studies look like since a germline gene intervention has to be done very early in life, and Alzheimer's doesn't present late, so he's going to need some pretty good long follow-up, but that's his goal. That, it seems to me, is a pretty avant-garde approach to the purpose of medicine. It's arguably illegal. Many people would think it's immoral. Jiankui thinks it's the next step. Compare him to Fyodor Urnov, who works at Berkeley and has been a pioneer in the development of CRISPR therapies.

John Lantos

He was one of the people who helped develop the treatment for sickle cell disease that has now been approved by the FDA for treatment of sickle cell. But Urnov condemned He and said interventions in the germline are completely inappropriate and irresponsible. They should be prohibited forever. But he also said current regulations for doing gene therapy and gene editing in competent adults are much too strict, and that there should be more of a right-to-try approach. That the FDA is getting in the way of innovation rather than encouraging innovation. That IRBs are doing the same thing and that in these calculations the mistake people are making is overweighting the risks of research and underweighting the risks of not developing new treatments for terrible diseases. He says that underweighting over here is leading to a lot of potentially preventable suffering if we could only do research faster. I'm not going to take a side on that debate, but just bring them up to say here are people who are pushing boundaries in what I would consider an avant-garde sort of way.

John Lantos

They're both determined to prove that traditional biomedicine can deliver miracle cures in exciting new ways. Or as another example of what you might consider avant-garde consider two, should I call them public health leaders? Anthony Fauci and Robert F. Kennedy Jr. Again, I'm not going to take a side on which of these is doing great stuff and which should go to jail, but I think it's fair to say that Americans themselves are divided on that very topic because people don't know who to trust or what the basis of that trust should be. It's one of the ways that the controversies in medicine, delivery of drugs, development of drugs, public health interventions has lost its foundational base in a widely held consensus. Some of the controversy comes from the fact that the new cures that people are developing tend to be extraordinarily expensive. I mentioned the sickle cell treatment that's been licensed by the FDA actually now two have. They cost $3 million per dose of the medication, and that doesn't count the two months in the hospital that people need to get them. So new treatments will probably cost $5 million per patient.

John Lantos

Alzheimer's drugs. Six million Americans have Alzheimer's. If only one-sixth of them got treated for Alzheimer's, it would cost Medicare $29 billion a year. That's almost two-thirds of all that Medicare spent on drugs in 2019. Or I'm not even going to get into this, but if people have been following the muscular dystrophy drugs, they're just an expensive scandal. So should we trust the experts? A lot of people say the FDA, the NIH, the journals tell us what's good for us at those prices. At the same time, millions of people can't afford their insulin or their EpiPen or their asthma inhalers and wait months to see a specialist. This is a system that is ripe for creative transformation. Trust the experts. Stop sounding like good advice and start sounding like a sales pitch. So I said we were going back to Walter Benjamin and what photography did for art. The other thing it did besides create this exuberant avant-garde creation of new art forms was it radically democratized art.

John Lantos

Suddenly with photography, everybody's an artist. You don't need to go to a portrait painter. You get out your Kodak, your Brownie, your phone. Everybody is an artist. I think that is part of what's going to happen from AI in medicine. And let me give an example from my own experience as a patient. A few years ago I had two MIs.

John Lantos

I was treated. The treatment was excellent. I have a bunch of stents. I'm on a bunch of meds. Post-op they found some arrhythmias and put me on a Holter. And the Holter monitor results were alarming. Multiple runs of V-tach. Before I show this next slide I just want to say to the cardiologist in the room, I'm doing okay now. But this was what my Holter looked like. Yeah.

John Lantos

And I consulted many cardiologists. Being a doctor I knew a lot of cardiologists. I called them all up. Everybody had somewhat different advice. They told me there were all sorts of options. You could get a pacemaker. You could get an implantable defibrillator. You could get medication. But they also said you don't fit into any of the categories that had been well studied. Your ejection fraction is still pretty good. You were not having symptoms even when you were having these runs of V-tach. And the head of Harvard's EP lab said, I've got to tell you buddy, you're in an evidence-free zone. Which is not always a reassuring place to be. They did an EP study. They found no identifiable source of the arrhythmias. Nevertheless, they decided to just ablate anyway once they were in there. They said it's probably coming from the scar tissue from the old MIs. We're going to ablate that. They implanted a [monitor name unclear] monitor that I'm wearing and have not had more runs of V-tach. Thank God. A follow-up echo showed I got a ventricular thrombus at the site of the ablation. Probably iatrogenic. So it's unclear whether I needed this in the first place or whether it got better.

John Lantos

I consulted ChatGPT because I wasn't getting the answers I wanted. And particularly because nobody would talk about prognosis. People said like, oh, take your meds. The studies, you're in an evidence-free zone. We don't know. It was hard to get an answer. So I said to ChatGPT, I want you to be two doctors. I want you to be an optimistic doctor and I want you to be a pessimistic doctor. I want you to give me the best case scenario for my prognosis. I want you to give me the worst case scenario. So that's what this was like. And one of the things in the democratization of doctor-patient relationships brought about by AI is everybody's going to be able to do this now. And this image was generated by AI. But what it imagines is you have sliders where you can say empathy, autonomy, paternalism. You can design your own doctor to be emotionally the sort of doctor you want. So the optimist said, you have moderate heart disease, not advanced heart failure.

John Lantos

Your EF's 45%, not in a high-risk category. With treatment, many patients with that lived for decades. Among 70-year-old men, [brief phrase unclear], many patients are alive at 80. The pessimist. You have ischemic heart disease with MI damage, reduced ejection fraction, ventricular scars, documented V-tach. Each increases your mortality risk. Your chance of 10-year survival is lower than age-matched peers without heart disease. Even with optimal management, your risk of death is substantially higher than average. Both of those are true. Right? If you go to a doctor, you're just going to get one of them. If you go to many doctors, you'll get many of them. But with AI, you can get them all at once. And you can fine-tune them. And you can ask follow-up questions. So it helped me understand the power of AI for patients. And led me to the belief that focusing on how this is going to work for doctors.

John Lantos

Like whether a better cardiogram reading machine or an AI-empowered stethoscope are going to be the transformational elements of this. Is the wrong focus. The real transformation is like the democratization of art. Everybody becomes an artist. I could ask as many clarifying questions as I wanted. I could ask them at 2 in the morning. When I said, what's the basis for that claim? ChatGPT gave me three citations. I could design my own choice architecture. That's the term of art for the way doctors frame and nudge. To get patients to make the decisions that they want. The one thing that patients have no control over. But with AI you do. You can ask for the optimist or the pessimist. It empowered me. And it enhanced my autonomy. I think there are two disruptions here that are likely. And making predictions about where AI going is, of course, fraught with uncertainty. But I think patients will increasingly demand the interventions that LLMs recommend.

John Lantos

And will not want their doctors or regulators to stand in their way. I think people will have the sort of experience that I had where I compare what I get from the LLM with what I'm getting from doctors. And say, I'm going to go with the LLM. And why should I then have to go to my doctor and say, ChatGPT says you should prescribe this. Why can't ChatGPT acting agentically just send the prescription in? I think more and more that is going to be what patients demand as they have this experience of comparing what they're getting from their human doctor and their phone. Second disruption. I think for doctors we are going to need to get used to and adapt to what it means to be the physician for someone who has an AI primary care doctor. And they are going to come to us with much more sophisticated information. This is, I mean, I think doctors started to face this when people talked about Dr. Google and people doing their searches.

John Lantos

But that was primitive technology compared to what we're seeing now. And I think people are going to start coming in with much more sophisticated questions. And we are going to have to learn to deal with it. So in my view, the future can be learned from the past and particularly from the sort of ideas that Walter Benjamin was talking about. The camera didn't kill the painter. It freed painting to experiment. And the question is, will we as doctors allow ourselves to be freed? If not, I fear we will go the way of the portrait painter. And perhaps we will deserve to. Thanks.

Recording

[Applause]