Robert Wachter, MD, A Giant Leap

Thank you, Amnut. It's a great pleasure to be with all of you today and look forward to talking about what I think is probably the most interesting thing I've seen in my very long career in healthcare, which is uh uh the advent of of the new artificial intelligence. So, I've called this uh when AI comes to healthcare, is this our Hemingway moment? And what I mean by uh by that is uh comes from I figured talking to the uh to the folks at the library. Um it's worth talking about books and so this is the 100th anniversary of The Sun Also Rises Hemingway's classic book uh about bull fighting and much more and in the book one of the characters goes bankrupt and another asks him how did you go bankrupt and the answer he says famously is I went bankrupt two ways gradually and then suddenly and I think this is our suddenly moment in healthcare when it comes to digital transformation you think about every other industry. The way you interact with your bank, the way you make uh you buy things, the way you make plane reservations, uh the way you make restaurant reservations, the way you park, uh all have been transformed by technology, and yet uh I'm taking care of patients today here at at UCSF. Yes, we record our information in an electronic health record and our X-rays now are on computers rather than on pieces of film. Uh but um medicine has been curiously untransformed by health care. Still much of the we're still the major users of fax machines in the uh in the world until recently. We carried beepers where even the drug dealers got rid of theirs. But I really believe that the new AI, generative AI, is is our hemiway moment where we you're going to see real transformation in the way your health care is delivered or if you're a clinician, the way you deliver healthcare. uh it's already starting but over the next few years. So that's kind of what I want to focus on and talk about why that's happening and what's good and what's bad about it. And uh as you'll hear I think it's more good than bad when it comes to healthcare. Uh so this is the book I wrote uh came out about six months ago. It's made a few bestseller lists. It's been quite gratifying. Um when you write a book you try to find prominent people in your field to write blurbs. This is a a physician named Eric Toppel, very famous physician in the United States. And Eric's blurb uh was he called it an engaging hype-free case for informed optimism. And I leave that with you because I think that is where I landed after two years of research and interviewing about 110 leaders in health care patients, clinicians, policy makers, doctors, nurses, uh former FDA commissioners, etc. I landed in a place of what I think is informed optimism. And I'm going to share that with you and hope that you share that as well. Um, a lot of people have told me that the book is very readable and lots of stories and interesting characters. And I've even had some people tell me they couldn't put the book down. Uh, one person could. This is my three-year-old granddaughter. >> Hold up the book. >> The book. >> We need We need the >> So, there we go. she was capable of putting the book down, but I think most other people have found it a pretty compelling story and really kind of amazing uh people, companies and questions, including, you know, big cosmic ethical questions about what happens uh when a field like healthcare goes from doing one way of work to another. Now, part of my optimism uh is because this the new AI really is quite good and can do a whole bunch of things that nothing that we had up until four years ago could do. But part of it really is the predicament that we find ourselves in in healthcare. And let me kind of describe that to you. This is the way I think about it. On November 29th, 2022, and u you may recall that is the day before chat GPT was released to the public. On that day, I didn't think that for example, I needed a better search engine than Google. In fact, I'm not all that creative. I couldn't imagine a better search engine than Google for it. I thought Google was search. You put things into that little box and it gave you a bunch of links and that was really great. And I'm old enough to remember before we had that how you did search. You went to the library or pulled out microfiches and and I thought that was sort of the epitome. I couldn't it was the height of search. Couldn't be any better than that. And then the next day of course GPT came out and I said actually this is better. Uh and Google thinks it's better when you do a Google search now. The first thing you get is an AI search from Germany rather German Gemini rather than all the links. So, I think a lot of industries were like that where I didn't think I needed AI to help me plan dinner or plan my trip to Italy. But when the new AI came out, I said, "Wow, this is better than what I had before." And I used it. Let's contrast that with healthcare where I don't know anybody uh who was sane who on November 29th uh 2022 said the health care system is perfect. I actually can't imagine a better healthare system than the one we have. I uh I I feel that even at you know the US news rankings came out last week. US UCSF was ranked as one of the top five hospitals in the country. Uh it is an amazing place with extraordinary people, incredible technology. And yet we all know that quality isn't what it needs to be. Safety isn't what it needs to be. Certainly hard to get in and see a doctor. It's not convenient. Uh and the costs are backbreaking. And so I don't think anybody thought the health care system was optimal. And let me tell you three of the areas where it's just very clear that without AI we couldn't solve the kinds of problems that we have and with AI we're already seeing uh them at least partly solved. And the three issues relate to the amount of information that we have flowing around the health care system and how difficult it is for both doctors and patients to manage it. So these three cases relate to uh what happens when you see the doctor and the doctor's asking you a bunch of questions. uh what happens when the doctor has to review your old medical record and then what happens when the doctor needs to uh figure out based on the latest medical literature what the right thing to do is. Let's take the your you go to see the doctor in the office or the emergency room and doctor's talking to you. Uh here's a picture that a seven-year-old girl drew now a little over a decade ago after her visit to the pediatrician. And you see her sitting on the table. Mom's next to her sister in the corner. And there in the far corner typing away back to the patient in the family is the doctor. This is what she recalled of her visit to the doctor. For those of us who are doctors, this is a very familiar uh picture. There's one thing the girl got wrong uh and that is the the smile on the doctor's face. I know of no physician, including me, who became happy once we had computers uh becoming a pretty expensive, grumpy data entry clerk. Uh but we once we had electronic health records, we had to fill out so much information and check so many boxes that you as a patient may have noticed that your doctor wasn't looking you in the eye but was looking down at the computer. Today at UCSF, we all use and have access to an AI scribe. Uh, so if you came and saw me, I would put my phone on between us, ask your permission to record the conversation, and it would create my documentation, which is not as simple as just creating a uh a uh transcript of our conversation. That actually would be pretty worthless. Uh, if you say, "I'm short of breath," and then 5 minutes later in the conversation you say, "I have chest pain." Those two things need to go together in the in my documentation. Uh, it's lovely to hear about your trip to France, but that doesn't go in my documentation. So the new AI tools are able to to put things in the right place in the right order, purge the things that are not particularly helpful. Uh and that is a tool now that's become wildly popular and as I say every doctor at UCSF now has access to that and we're using them all the time and patients are noticing too that we're actually looking them in the eye which we couldn't do until recently. So that's one problem that AI has helped us tackle documenting the visit. before I see you, there's another documentation problem, which is I've got to review your chart. Review your medications, review your history, your medical history, your surgical history, um, things like that. And here's a study that came out a year ago that said one out of five patients has a medical record longer than Moby Dick. And so, if you've forgotten your classics, Moby Dick is 600 pages long. So the idea that in the two or three minutes that I have to see a patient, I'm going to be able to review 600 pages of data and summarize it and pick out all the relevant facts that are that I need to understand to make sense of what's going on with you today. Uh that is uh impossible. If I had an hour maybe, but I don't. And so whereas today we can use an AI tool that will summarize all 600 pages link to where it's finding information. and it's it's not perfect, but it's pretty darn good and it's better than I am when I try to do the same thing. So, that's a second problem, reviewing the the data in your medical record. And the third problem is reviewing the data in the medical literature. And I will confess to you that of the 8,000 medical articles published per day, I get to read maybe five of them, maybe 10 of them, uh, but impossible to keep up. And of course, AI can do that as well. And most of us are using AI tools where we'll put in information about a patient and it says you know the latest literature says the right treatment for this is so and so or the right test to do for uh this diagnosis is something something else. Uh so these are three areas where I'd say they are simply not fixable without AI and the status quo of what we all did before AI was highly highly imperfect. And that's kind of part of the message about why I'm optimistic about AI and health care. I'm I'm probably as worried about it as everybody else when it comes to uh the to climate, whether my kids are going to have a job, etc., etc. I think there are bioteterrorism. There are real concerns that are legitimate. Uh but in healthcare, I think uh I am net optimistic in part because these are the kinds of problems that really seem to me to be not not fixable uh without AI. Now, the public agrees with me. This is a survey from Gallup uh from last year that showed that um people are really quite skeptical about AI in pretty much everything. And this is last year. I'm pretty sure this would be worse today. Feels like the only thing that the right and the left agree on is to be against data centers and to be anti-AI. Uh and you see that most people were worried about false information, social connections, jobs, national security, etc. The only area where people were generally positive is medical diagnosis and treatment. And I do think it's because they like me perceive the health care system to not really being a be able to deliver uh despite lots of smart doctors and nurses and lots of wonderful places like my own. Uh but it simply there's not the capacity to deliver what patients need without AI. So how do I use AI? I'm on clinical service now. As soon as I'm done here, I'm going to go out and see a bunch of patients. uh here at Parnasses at UCSF and I would say today I used it seven or eight times. And how do I use it? I use it in what we physicians call a curbside consult. And what that means is if you're my patient in the hospital and you really need to see a cardiologist because you know you need a procedure. I will call a cardiologist. The cardiologist will come and see you, examine you, talk to you, review your chart, write a note, etc. Uh, but for every one of those I do, there are probably five times where I, God, I'm I'm a general doctor, so I take care of patients with heart and lung and cancer problems and infectious problems and kidney problems. And I'd say 10 times a day, I have a patient where I'm pretty sure I know what to do, but I'm not 100% sure. And it would be nice if I ran into my favorite uh hematologist in the hallway or or infectious disease doctor. When I do, we do this thing called a curbside consult, which means I say, "Emma, you know, I have a I have a case I'd like to run by you. Can you help me with it?" And everybody's very collegial here, and they generally will, and I'll briefly summarize the case, and they'll briefly give me a what they think is going on without doing a formal consultation. Those sorts of curbside consults, which depended on something kind of silly, which is me running into somebody in the hallway. Uh, not quite, as they say in Silicon Valley, scalable. I now use an AI tool. Sometimes I'll use a general tool like like like GPT or Claude uh the same kind you might use. I'd say most of the time I use a tool that's only available to physicians. It's called Open Evidence. It's really GPT but built for doctors that uses the medical literature. Uh but here's an example of me using just GPT to talk to it with a clinical case and get an answer that I think is as good as the answer I would have gotten if I ran into my favorite specialist. So, this is a clip that'll take a minute or two. Hi, GBT. I'm a doctor. I'm taking care of a 28-year-old woman who comes in with a couple of months of unintentional weight loss, maybe 15, 20 pounds. Uh, she's also tremulous. Uh, has noted some cardiac palpitations and has had some diarrhea. Uh, can you briefly tell me what you think the likeliest diagnosis is? >> Absolutely. So with those symptoms, the unintentional weight loss, the tremulousness, the palpitations, and the diarrhea, hypothyroidism really does stand out as the most likely cause. It's kind of that classic picture of an overactive thyroid revving up the metabolism and causing all those symptoms. So that would be the top thing to check out. >> Any other diagnoses that are kind of can't miss diagnoses? Ones that if I don't get it right, could really harm or maybe even kill the patient? >> Yeah, definitely. So besides hypothyroidism which is the most likely you want to keep an eye out for any signs that this could be progressing to something like a thyroid storm as we talked about earlier that's a kind of can't miss urgent scenario if it were to occur. Another thing to consider is any kind of cardiac arhythmia that might be driven by hypothyroidism like atrial fibrillation that can sometimes lead to complications like heart failure or even stroke if it's not recognized. So you want to keep an eye out for that. And very rarely you think about things like a fiochromocytoona which can also cause palpitations and weight loss but that's much rarer. In short hypothyroidism is top of the list but you just want to make sure you're not missing any severe cardiac issues or a thyroid storm scenario. >> Great. Uh can you tell me a short poem about hyperyroidism? >> Absolutely. I'd love to. So here's a little poem for you. When the thyroid is dancing a bit too fast and the heart races on like a drum. The hands start to tremble, the weight can drop fast in the whirlwind of hormones that come. So there you go. A little poetic nod to hypothyroidism. [clears throat] >> So there you go. That's uh fairly typical use. That's an answer that I think is pretty much as good as the one I would have gotten from my favorite endocrinologist. And uh of course the little poem at the end is is is a party trick, but um it is pretty darn good and much much better than it was uh when I first played with it about four years ago. Now let's not sugar coat it. It it's not perfect. And in the first year of generative AI, uh there were these cases that we all came to understand as quote hallucinations where the AI not only got it wrong but kind of BSed us and made up a whole scenario that was sounded believable, but it turned out was completely fabricating it. And uh here are a couple of examples. If you said, "I'm having trouble. My cheese is sliding off my pizza." Uh, one of the AI tools said you should put glue on your pizza to keep the cheese on. If uh, to their credit, they did say non-toxic glue, but still not a great idea, nor a great idea uh, to recommend eating at least one small rock per day. So, these were the sort of things that came out in the first year where I think it was logical to say, "Oh, this is kind of fun for, you know, planning my my trip to Yoseite, but maybe not for a field like healthcare where we can hurt or kill people if we get it wrong." But I would say over the last four years, these tools have gotten massively better. And Ethan Mollik, who is a brilliant uh professor at the Wharton School at Penn, this is his book called Co- Intelligence. He's not got another book coming out in a month or two. Um, and he's really the most thoughtful writer about AI and the workplace and AI and education. He doesn't write about healthcare at all. Um, but one of Ethan's rules, which I think is a fair rule, is assume the AI you're using today is the worst you'll ever use. I think that's a reasonable assumption. And the kinds of hallucinations that we saw in year one, uh, in 2022, 2023, you rarely see anymore. And I have to say, when I use a tool like Open Evidence for a curbside consult, it gives me the right answer or answers I find helpful more than 95% of the time. I'd say it's really quite impressive. Now, as AI gets better, uh there's a final challenge. And since everybody's in San Francisco, this will be a familiar challenge. Um I remember in the early days when people said, well, you know, we can't use AI in healthcare because healthc care is really high stakes and if we get it wrong, we can kill someone. Well, you know, I I don't know about you, I take a Whimo about once a week. In fact, my wife and I ditched our second car a year ago and we just take away when we need to get somewhere. Uh, and part of the reason we do that is it is demonstrabably now there's no question the research says it's safer than if I drove the same car myself. Any time people say, well, AI is no, you know, we can't use in healthcare. The stakes are too high. I say like try making a left turn across to Viziderero during rush hour. Like the stakes are high and if you get it wrong, you can kill kill somebody. And yet many of us are very comfortable taking these cars because there's no question that it's safer than a car with a driver. I think I think my latest data from Whimo is more than 250 million miles and not a fatality caused by the car yet, which is an remarkable safety record. And yet, as people remember, a little less than a year ago, Whimo ran over a poor little KitKat in in the Mission District. It had the misfortune of running over the most famous cat in San Francisco. Here's a shrine to KitKat on 16th Street. And um it made international headlines. So it's one of the challenges of AI and healthc care is even if it's really good and really helping us and maybe saving lives, if it causes if if it causes harm at any point, it will be an international story. I don't know of a case yet where AI has killed anybody in healthcare, but it's inevitably going to. uh the stakes are too high particularly as we use it in riskier situations. And so Whimo did not go out of business in part because its safety record is so good. But this is the sort of thing that happens when uh a technology enters a high stakes field. And even if it has an enviable safety record and is safer than the alternative, uh when something bad happens, you'll hear about it. There will be lawsuits. There'll be newspaper articles. And uh uh as Joe Biden liked to say, don't compare me to the almighty, compare me to the alternative. And yet when it comes to a technology, we tend to call compare to the almighty, which is sort of what happened here. Um here's a problem which is of human vigilance. And here's here's the way I think about it. If AI was correct half the time, uh let's say drafting my note uh from the visit with an AI scribe or reviewing summarizing a chart or even suggesting a diagnosis, uh it would be completely worthless because it's half right half the times wrong half the time. I couldn't trust it at all. I'd have to review it carefully myself. Why even bother using it? If AO was right 100% of the time, that would be pretty great. And actually, you would not want a so-called human in the loop because all the human could do would be to degrade the performance if it's starting out perfect. It all it can only go down. I'm not sure what any of us are doing for a living if that happens, but that's a different a different problem. Uh but the the issue we find in healthcare today is the AI is right often enough to be useful, but not perfect enough to be entirely trusted. So, we're going to set up systems where humans will be asked to be the safety check. Meaning, the AI drafts the note, but I review it. The AI summarizes the chart, but I review it. The AI suggests a diagnosis, but I, as the doctor, I'm the final arbiter. That sounds better than it is for a few reasons. One is the problem of human vigilance. There are a lot of things humans aren't very good at, but I'd say pretty high on the list is remaining awake and alert if an AI or any technology tool has been right 20 times in a row. You know, are you going to be carefully reviewing number 21 or are you asleep at the switch and doctors are no more likely to be uh awake than than any other person is. So it this this this system that we are typically setting up which is the doctor is the final arbiter or the nurse is the final arbiter is a little bit dicey because humans are not good at remaining perpetually vigilant if they've come to trust a tool. The second reason it's not perfectly failsafe is the problem of of deskkilling a problem we've seen in other industries. This is the crash of Air France 447 off the coast of Brazil in 2009 where a little gizmo on the outside of the plane froze. The pilots were flying through a storm at night, couldn't see anything, and they and they had they lost all their instruments and all their technology help. And they made all the wrong moves and flew basically an intact airliner into the ocean, killing about 250 people. And I remember interviewing Captain Sullenberger, who famously landed on the Hudson, ironically, the same year, 2009. And Sully said to me, I said, "What do you think happened with Air France 447?" and he said the pilots were flying a plane they weren't familiar with. And so this issue of deskkilling is is going to come to medicine. This is the first study bringing it to light. This is a wonderful study from last year from Europe where colonoscopists, these are gastronurologists who stick that long tube up your rear to look for precancerous lesions in your colon, were given access to an AI colonoscopy tool that helpfully puts a little box around lesions that look like they might turn into cancer. That's great. They used it for three months. They thought it was really spiffy. They and it actually improved their performance. Then they went back to their old colonoscopes that didn't have the green boxes and their performance their their the frequency with which they found precancerous lesions went down significantly. And you might say with these newbies now they these were the these gastronologists had been doing this procedure colonoscopy for on average 10 years. So very experienced people using this tool for three months became less good at the thing over the course of the next six months. Uh that wouldn't be terrible if they had access to the tool every place they practice, but the tool is not universally available. There's certain desilling that's fine. I no longer know how to read a map, but that's fine. Uh so it may be that these tools become ubiquitous. We all have access to them and if we lose the skill we used to have, that's no big deal other than nostalgia. But for the time being a lot of AI tools are available in some places not available in others may not be perfectly reliable. So this is another problem with humans being the final arbiter of a imperfect but very good AI output. Kind of the final issue I want to talk about here is is is the issue of who's using the AI. I've talked about what happens when I use it. It's not that I'm so spectacular, but I've been a doctor for 40 years and practice in a very good place. Um, when I like the the uh the video I showed you about hypothyroidism, you may notice that what I put into it was a very uh sort of um focused sample of the data that I might have had about a patient. When I say this is a third 32-y old woman who comes in with fever, shakiness, sweatiness, and diarrhea, it may be that I have a hundred facts about the patient, but I've chosen to put four in. The reason I do is I know a lot about medicine and those are the four that strike me as being the most relevant. And when I do that, the AI tool gives me often the right usually the right answer. When patients are interacting with G with AI tools, GPT or Claude or Gemini, let's say, uh they don't know and no reason they would know which inputs are important or when they get outputs, which ones seem right, which ones seem wrong. And so we are seeing examples of that. This is a study that came out last year from Oxford where patients were given uh prompts to and then told to interact with the with AI and the AI got it wrong a little bit more than half the time. When you just put the prompt that the patient had been given into the AI, the AI got it right 90% of the time. And it was that the patients didn't know which parts of the prompts were important and often put in the wrong data. Whereas trained professionals, if they're any good, they know what to put in and they know how to interpret the output. I'm actually going to go past this because I want to be sure we have time for questions. But the point is that we think about this a lot in medical education because when my daughter started medical school here, she was a novice and knew, you know, didn't know any more than anyone else about what to put in. By the time she finished four years later, she was an expert and knew and one of the things we teach uh quite rigorously is how do you what is the patient data? What are the questions you ask and what's what are the relevant facts that we might want to use if we were interacting with AI or for that matter speaking to another doctor. So the tools I think can and need to get better. They need to act a little bit more like physicians do when a patient has a complaint. First of all, they need to have data about your past history in order to give context, but they they they shouldn't be giving the answer right away. They need to know what questions they need to ask you in order to to have the data they need to to uh to be effective. How should patients use AI? I think it's fine to use it. It's better than nothing. I think using it before a doctor visit, asking the AI what questions you should ask your doctor or what questions you should be prepared to answer. after a visit, putting your doctor's note in, lab tests, and other results back into AI and ask for an explanation in plain English if English is your language. If you wake up and you have a sore throat or a headache, I think it's reasonable to put provide detailed symptoms, but how severe it is, when it started, uh what's made it better, what's made it worse, and also what other conditions you have. Do you have diabetes or hypertension or or uh have you had a heart attack? And what medications you're on? And of course, if you use some of the tools like GPT Health, it asks you to load in that information from your electronic record, you're going to have to decide is that are you are you willing to accept what I think is a very small but non-zero risk on privacy uh by sharing your record with with a company like uh like OpenAI or any of the other uh AI companies. Keep your eyes open for better tools and I think ones that act more like a doctor. And the tools will be better if you give them your health information. As I said, you're going to decide have to decide whether you're willing to do that. When you give your health information to me at UCSF and I put it into our electronic health record, which is built by a Wisconsin company called Epic, uh we all operate under a set of rules called HIPPA, you may have heard of, which provides very very strict penalties if any of your personally identifiable data leaks out. When you give your data to one of the AI companies, uh they don't have to operate under that uh that set of rules. So you have to believe them when they say you they're not sharing your data with anyone. I generally do and actually did put my health data into GPT. Wanted to see kind of what it what it did but it was with you have to do it with your eyes open. Few final thoughts. Will AI replace doctors? I think it turns out that even in fields that we thought would be easy to replace like radiology and pathology which after all are really just today about reading a collection of digital dots and saying that looks like a certain pattern. Uh, it's far hotter than tech folks believe and famously Jeffrey Hinton who's won the Nobel Prize for basically founding the new AI in 2016 gave a speech where he said I think we should stop training any new radiologists. It's completely obvious that within 5 years AI is going to do better than radiologists. Uh, Hinton was very smart about a whole lot of things but on this one he was not. So by his prediction, we would have not needed any radiologists by 2021. And I can tell you at UCSF today, we can't hire radiologists fast enough to keep up with the demand. And and you know, and yet we're all taking or many of us are taking cars with no drivers. So this turns out to be harder than anybody thought. Uh I asked one of our AI radiology experts here at UCSF, what did Hinton get wrong? And he said uh basically Hinton thought that reading an ex reading a cat scan or an MRI was pretty much the same as differentiated uh differentiating a cocker spananiel from a cinnamon danish on Google images. And it turns out it's massively harder than that because of human variation because the context is important. Uh will will AI ultimately help read uh X-rays? Absolutely. Will it replace radiologists? I don't think completely, but it will do a whole lot of basic readings. I think it will be able to tell whether a CAT scan is completely normal. Uh, but I think for for the foreseeable future, there still will be a radiologist doing the final read. Uh, and I believe ultimately that the the mass of unmet needs in the health care system will mean that AI will not displace many doctors or nurses for the foreseeable future. Does that mean nobody's losing their job in healthcare? I don't think so. Uh here we have probably over a thousand people. UCSF has about 30,000 employees. I'd say more than a thousand people are involved in taking the medical notes and and and massaging them and sending them off to an insurance company. Uh using the right codes um understanding that in this insurance company's rules are different than that ones. Do we need a thousand people if we have really robust AI? Probably not. Uh, part of what those people do is generate letters to the insurance company begging for them to allow us to do your scan or give you a certain drug. Uh, what's called a prior authorization. Can AI do a lot of that? Yes. Uh, also, if there were lots of unemployed doctors or or nurses, uh, we would be pushing back pretty hard. Uh, over a 10 to 20 year time horizon, who knows? I mean, I think it's hard to know in any profession what's going to happen, but I suspect there will still be uh doctors even even 10 to 15 years from now. And as I said, today in San Francisco, we're many of us are taking driverless cars and we cannot hire radiologists fast enough because we're doing more and more and more scans. And even if the AI makes us 50% more productive, uh that's barely enough to keep up with the scans. So, we're at I want to be sure we have time for questions. I'm going to end here by reading a and this will be spoiler alert for if you haven't read my book. I'm going to read the last couple of pages because one of the things I tried to do in the book was think hard about you know what do I do as a physician? What can AI do? And will AI replace me uh and try to do that without my ego being too deeply involved in the decision-m because I can imagine in a world where you know I I mean I don't see a travel agent anymore. or I make my own travel arrangements. I do much of my own finances. And you could see a world where if AI could do a whole lot of what doctors do, patients would be happier. It'd be less expensive. It might be more convenient. So what I did to try to answer that question was I I did what I'm doing now, which is I'm on a clinical service and I tried for about 10 days to use AI everywhere and really think hard about, you know, the AI we have today, but what might it be like in a few years as it gets better and better. And so this is how I reflected on that experience. It would take about two minutes and then we'll be open for questions. Toward the end of my time in the wards, I cared for a 22year-old woman with severe autism who was non-verbal at baseline. She was brought to the ER by her mother for fever, cough, and lethargy. In her case, the diagnosis of pneumonia was straightforward, as was the choice of treatment, IV fluids, and standard antibiotics. I had no doubt that AI could have easily replaced my clinical decision-making. Still, and and perhaps I'm being selfish, I treasure having been part of this young woman's care. As she emerged from her near comeomaos state, her mother's face brightened as she proudly told us that her daughter was back to her quote sassy self. Thinking of my own experience as a parent, I tried to put myself in this mom's place. I was aed by her resilience, her ability to find delight in loving a child who would never speak a word, for whom life's conventional milestones would always be out of reach. Lost in these thoughts, I turned toward my patient, who was only a bit younger than my team of residents and students standing by her bedside. Just then, the patient flashed us a radiant smile and blew kisses to me and my team. As I hope I've made clear, I am convinced that AI's breathtaking and rapidly expanding capabilities will transform healthcare in ways we can barely imagine today. Bringing enormous benefits to patients and clinicians and turning our health care system into one that for once functions reasonably well by handling myriad administrative chores, allowing patients to accomplish many health care tasks independently, and permitting clinicians to be more capable versions of their professional selves. AI is poised to make our system better, safer, more convenient, and less expensive. But I'm equally convinced that particularly for patients with acute and complex chronic conditions, there will always be a need for a human guide. Someone with not only deep medical knowledge and refined clinical judgment, but also the emotional intelligence to recognize and address their unspoken fears. The leadership skills to orchestrate their care across diverse teams. the patience and wisdom to navigate the inherent uncertainties of medicine and healthc care's bureaucratic culde-sacs, the uniquely human capacity for deep compassion that transcends both the practical and the algorithmic and someone to be on the receiving end of a speechless patients kisses. They will need a doctor. With that, let me stop. Thank you for your attention and I look forward to hearing uh hearing any questions or comments you have. Thank you so much Dr. Water. Um we have a few questions from the audience. Um the first one is um over the years the advent of different tologies and servicebased pools have allowed us to save time and consolidate patient reports. But what happens when this bubble bursts and some of these pop-up companies inevitably fail? What happens to the patient data if the company that goes under and holds the data storage are no longer available? How are we protecting patient data and remaining HIPPA compliant? >> Yeah, it's a great question. I remember uh when I used back in the day 15 years ago, I used clear, you know, the the service that helps uh expedite your your uh your your trip through the airport. And uh and they had my fingerprints and my iris print. and then I got a notice that clear went out of business and I had no idea and they didn't say what happened to your your data. So I think it is a concern. I think when you are taking your data and allowing it and at least in the case of allowing your data to go uh from your Apple Watch or from your hospitals or doctor's electronic health record to a third party company, you're going to generally be asked to click a little box that says you understand that it's going off to this other company and the company's promising to do XYZ. When that company goes out of business, I don't know if they have any obligation to, you know, to burn up your data. uh or could they potentially sell your data to someone else that that isn't living by the commitment that that company had. So I think when you take your data and move it around outside of the health care system, it's a little bit of buyer beware. Um when I and my wife also decided to take our data out of not out of it's still in Epic which is you know our data from UCSF and move it into GPT and give it to OpenAI there was a little box that says you you know we promise not to sell your data use your data for training center and we both said you know do we believe them and we said you know the risk they the the scrutiny that they are under on privacy is pretty high and thought that the risk of that was out you know out was less than the potential benefits. Where this goes over time, what I hope happens is you don't need to do that. You don't need to take your data out of your health care systems electronic health record and move it to a company's uh AI, but your your health care systems electronic health record. Most people are in systems that have Epic, which means you have my chart provides the same kind of service. So it stays within the database of UCSF which I don't think is going out of business and VPIC which is not going out of business which I think is probably ultimately safer and ultimately probably better than having your data move around to a whole lot of different companies whose future I think you you we don't it's impossible to predict. >> Thank you for that Dr. Walter. Um, second question is, is the AI consult built to the patient or identified as a line item on a hospital bill? So, that's a that's a great question. The AI consult that I'm using and when I pick up my phone and use GPT or use open evidence. I don't bill anybody for it. It's all part of my service. Now, it probably gets embedded when I bill for a patient visit. One of the things I'm asked is how much time the visit took. And so if that took me two minutes, I sort of embed that in the overall amount of time I took. But no, it's not. We wouldn't charge anybody for uh or send a bill off for an AI consult, nor do we do it for uh a curbside consult, which sometimes makes the consultants a little bit unhappy because they're taking time and their expertise is being used. But we all recognize that to provide the best care, we all need to do this. So everybody kind of agrees. uh when we do electronic consults. So if you know if if at UCSF um you're being seen by a doctor, let's say your primary care doctor and they send a note off to a specialist to get a brief consultation, the specialist doesn't see you. If the specialist does spend any significant amount of time, then there is a billing code for that. Uh whether ultimately we will be able to bill for AI services, I think is an open question. right now as certain AI tools get approval let's say from the FDA uh Medicare and other payers are looking at some of them and because often the companies are asking for an additional payment uh if if for example we bring in a tool that helps us read your mammogram and we do it because it will do a better job than the radiologists doing it uh doing it themselves. um right now actually of as of 6 months ago there was no additional payment for that. So we have to decide is it worth it because it's going to so improve the way we do mamograms or maybe improve the efficiency of our radiologists. But over time I think you're going to see those companies petitioning the payers to get an extra payment for the use of AI. Right now most of the AI we're using is we're just using it to to take better care of patients and there's not a bill being sent off to someone for the AI. >> Got it. Thank you. Um, next question is a short one. At four minutes in, you mentioned UCSF was ranked at what number? >> Yeah, the US News and World Report comes out every year with its ranking of America's best hospitals. They they say that it's just a overall ranking of the top 20 and they don't rank them, but we've actually seen the numbers and we we came out as fifth uh this year and I think tied for the best hospital in California. Okay. and uh number one in the country in two different specialties both of which live in my department I'm proud to say one is geriatrics and one is lung medicine and in the top 10 for several other departments including uh endocrinology rheumatology um and gastronurology so very proud this is a really terrific place >> yeah wonderful okay next question how do we as practitioners ensure that AI assisted tools do not create an even greater divide in healthc care disparities. >> Oh, it's a wonderful question. You know, if I was on my high school debating team and I think if I was asked to take the side, you know, where you were asked to take an argument one or the other, I think if I was asked to take the side AI is going to worsen disparities or AI is going to improve uh disparities, I would take the improve disparity side. And the reason for that is, you know, try to find a primary care doctor in, you know, in the Bay Area. Try to find and, you know, and one that isn't incredibly rushed, uh, or one that's going to see you if you have medical. Uh, you know, we see everybody, including patients medical. A lot of clinicians in the Bay Area don't won't. If you're in a rural area, you may be seen by a general doctor, but let's say you need to see a cardiologist or oncologist. I think the ability of AI to scale, make available specialist knowledge or make much more convenient uh the ability to see someone who maybe they're maybe it's a nurse practitioner, but he or she is using AI and it's allowing them to to do things and know things they couldn't have known before. I think that's going to improve disparities because I think the alternative, again, this is the Biden quote to compare me to the almighty, compare me to the alternative. The alternative in many cases is nothing. The the alternative in many cases is, you know, it's impossible to get in and see a doctor and, you know, unless you go to the emergency room and you might have to wait many hours to see someone. And I think that's the way primary care is going to go over time. I think that a lot of basic primary care um you know getting your vaccines which by the way I still believe in treating your cholesterol treating your blood pressure uh even maybe even your diabetes are are algorithmic enough that I think for many patients they will get care that they don't have access to today via AI tools um and that will free up the primary care doctors that do exist and there's a massive shortage to see the patients that really need them, that have very complex problems, have multiple problems. Um, and I, you know, we're beginning to see movements toward that. The state of Utah just gave permission to a San Francisco based AI company to refill prescriptions. Um, not and not every medicine, but but medicines that are relatively safe. I think that's healthy. I'd like to, you know, they're studying it. We'll see if it works. But I think for many patients, you know, to go and see the doctor to get a refill and have to take half a day off from work to do that instead and wait in the waiting room, you know, no patient really wants to do that. No doctor really wants to do that. So, I think we're going to see AI fill some of the gaps that are responsible for some of the disparities. You know, if it turns out that that only the richest hospitals have really good AI and poor hospitals don't, then you could see a a world where it worsens disparities. But I think it's more likely to be that everybody now has access to relatively affordable AI tools that allow for care that they couldn't achieve today and therefore it will improve equity across the system. >> Thank you so much Bob. Okay, next question. I think there is no replacement for human compassion which a doctor provides. I think the organizations for doctors and nurses employment should be educating the health professionals of having no fear of AI plus educating insurance um companies too. Don't you agree? >> I I do but I try to understand my own biases. I mean I have to say after 40 years of medical practice to say that AI could replace me is really you know saddens me. Um and uh and you know my daughter and son-in-law uh in at UCSF are both young doctors and they're looking at you know a career over the next 40 years. There are now studies that say when patients are interacting with either AI or a physician um through virtual systems and don't know which is which. They often feel like the compassion and empathy they got from the AI was better than that of the doctors. Now the AI has no compassion and empathy. It's a machine. And yet, um, uh, I think as these systems get better, I think that we will have to see. I think a younger generation may not expect that they see a doctor and may say, "Are you kidding me? Dad, mom, you used to go to this office and sit there and wait and read the newspaper while you waited 40 minutes to see a doctor for a 15-minute visit. They So, I think we're going to have to prove that. You know, I believe and hope and you know, as I go out this 10 minutes, I'm going to go see a whole bunch of very, very sick patients. Might have to tell a patient they have cancer. Might have to tell a patient that um you know, that they need to go there or talk about going on hospice. Do we want that to be a bot? Absolutely not. I think that that I will do that better and and and understand the human predicament and connect with a patient and hold their hands in ways that AI cannot do. But the the issue may be that that you know for a whole lot of medicine that we thought you had to see a person u that AI is able to replicate it maybe not just as well but also much less expensively much more conveniently and we'll have to sort of see patients you know I think lots of people do their taxes with Turboax and people but there are still accountants for people that really need it can afford it have complex needs uh every now and then I call a travel agent cuz I'm doing something really complicated. Now, medicine is not the same as your taxes or or planning a trip. But I think that we've got to be open to the possibility that AI may be able to replace some of the human connection that we have had while there are other parts of this that I think really are at least to me feel like they're they are indisputably human and I personally would not want an AI to tell me that I had diabetes or kidney failure or or cancer. And I would not want to be uh I think I think doctors do that better. I hope doctors do that better than than AI. So, we're going to have to see how this shakes out. Um, and some of the general skepticism about AI will of course influence this uh as people are more and more nervous about AI. But I think for young physicians, they sometimes say, you know, what's going to happen to my job? I think their jobs are safe for a very long time. And if physicians jobs go away, that means the lawyers and the accountants and the journalists and everybody else's jobs have gone away five or 10 years earlier. So, I think it's still an awfully good profession to go into. Thank you so much, Bob. Okay, next question. Do you think that AI would change the scope of practice for providers like physician assistants or nurse practitioners? >> I think that's a great question and in the book I spent a fair amount of time thinking about the history of NPs and PAs or allied health professionals. Um because in many ways in the old days there were a whole lot of things that we said, oh, a doctor needs to do that. you know only someone who's been trained like I trained you know four years of medical school and three years of residency and I did a fellowship and all that that's what you need to have this thing done and then in many parts of medicine we realize there's just too much work for if we say nobody else can do this but a doctor and the system does not have the capacity and we then began to ask are there other professionals who have less training and are less expensive and may not know quite as much but are very good and know a lot about certain things who could help and do part of this job. And when we did and those are NPs and PAs and I remember when we first started talking about NPs here at UCSF and they came to the credentials committee and were asking for things you know that that today seemed almost silly but it was like oh no no we can't have someone who's a not doctor doing that thing and today they do that you know times 10 we become much more comfortable that that someone who's trained that way and the training's gotten better and is experienced and you know I when I go into the intensive care unit probably a quarter or third of the the clinicians are NPs or PAs and are quite good. So I think the analogy is quite important because what it says is even as doctors sort of think and some of this is our ego and some of this is our guilds that you know oh no that this has to be a doctor task. Um we with NPs and PAs we we already cr we already crossed that Rubicon and we said all right this is going to be another person not a technology person who can do this piece of what I did and now we've got to think about parsing the work and saying here's a piece that I think this other person could do less expensively they're more available here's a piece that we really do need a doctor to do and we have that up and down the chain it's not just doctors versus NPSAs but also generalists like me versus a specialist And I think one of the things AI will do is now that decision of can we offload certain tasks maybe not to an NP or PA but to AI or to an NP or a PA with AI who now can do some things they couldn't have done before because they now have knowledge in their pocket or on their computer that they didn't have before. And the same thing analogously might be me when I'm doing those curbside consults. I am basically practicing the term we use is practicing at the top of their license. I'm actually practicing above the top of my license, meaning I now have access to specialty level knowledge. It doesn't make me an endocrinologist or oncologist, but it does mean some of the patients that previously I would have had to send to that specialist, now I'm pretty comfortable that I can manage that patient with the help of AI. So, it sort of rejiggers the entire system up and down the chain. patients being able to do some things themselves that they used to have to see somebody NPs and PAs doing things that they used to have to see a doctor doctor a generalist doctor seeing doing some things that they used to have to see a specialist and I think that's exciting I think that's going to be net positive but you know anytime you take a big complicated system that's also 20% of our gross domestic product and rejigger it like that you're going to have a whole lot of turmoil and and a lot of you know moaning by various professions worried about this and the regulator s will have a vote as well and the malpractice system will have a vote as well. You know, doctor, you know, shouldn't have you have gotten a real life cardiologist rather than just consulting AI. All that's going to happen and that will have some influence on the way this goes. >> Thank you so much, Bob. And that was the last question from the audience. Um with that, I want to thank you um Dr. Walter for sharing your time and expertise with us. um your thoughtful perspective on how AI is transforming along with your message of cautious optimism has given us much to consider. We're grateful for the insight you have shared um from your research and from your conversations with more than 100 experts. Um I also like to thank everyone for joining us this afternoon. Your questions and participation make for a wonderful discussion. We hope today's conversation gave you new insights into the opportunities and challenges AI presents for the future of healthcare. Um we'll be sending out a quick survey along with today's um presentation recording. Um your feedback really matters to us. So we love it if you could take a minute to share your thoughts and help us um make future session even better. So again, thank you everyone. We hope to see you at another library program soon. Um hope you have a wonderful rest of your day. Bye-bye now.