A Doctor's Guide to AI Medical Advice

0:00We talked about AI and the war at the top of the show. Now we'll ask how much can AI give you medical advice? Let's stipulate right from the start that chat GPT, for example, should not be your doctor. Do not forego a trip to the doctor's office because you believe chat GPT got your diagnosis right. That's not even medical advice. It's just common sense. But we'll talk to an actual doctor now who has some actual best practices, ways he says you can use AI medically, but he also has warnings. If you are using generative AI as a component of managing your overhealth or are curious about it. Dr. Adam Rodman, general internist and medical educator at Beth Israel Deaconess Medical Center in Boston, where he directs AI programs for the Carl J. Shapiro Center for Education

1:03and Research, and he's an assistant professor at Harvard Medical School. Dr. Rodman is both a practicing physician and an AI researcher, and maybe you saw his New York Times essay, a doctor's guide to using AI for better health. And by the way, it's not just patients. Part of the article is about how doctors should and shouldn't use AI, and yes, they are using it. Hi, Dr. Rodman. Welcome to WNYC. Thank you so much for having me, Brian. And listeners, we have time to take a few phone calls or questions via text. How are you using AI chat bots or generative AI for managing your health? Any successes or cautionary tales if you're a patient, or if you're a physician, call and tell us a quick story or ask Dr. Rodman a question, 212-433-WNYC-433-9692. All right, let's jump right in. Your first best practice in the article is use AI to enhance but not replace your medical

2:07appointments. People using it to avoid going to the doctor at all, and when is that risky? Yeah, so fantastic question. I know that there are people who are using it to avoid going to the doctor. Almost by definition, I don't see those people because they're not coming to see me. In my world, most of my patients are using it to help prepare for their visits, get additional information, or just understand more about their health. But there definitely are people out there who are using chat bots in lieu of going to the doctor. So that's one of the things that I really caution against. The next best practice for patients, because that one is kind of obvious, is about figuring out effective ways to describe symptoms. This one was really interesting to me. You're basically giving people advice on how to ask the most useful questions of a chat bot. Right? Exactly. So one of the, there's actually been a couple recent studies about many of the limitations which are very real in actual people using chat bots for their own health.

3:12And one of the big reasons is that large language models AI needs a lot of context. And because of that, when a doctor uses it, a doctor who has gone to many years of medical school and residency to understand patient context, it can have better performance. Patients obviously know you know the most about your own health from everybody than anybody else. You may not know what questions help differentiate different types of conditions. So my advice to my patients is if you want to use, you certainly don't have to, but if you want to use an AI system to better understand your health, have it interview you like a doctor? Have it ask you questions on what might be going on and have it interrogate you because it actually has in all of its training knowledge, it has encoded how doctors ask questions. How do you even do that? How do you ask a chat bot to ask you a question? Yeah, so this is what I would do. I would start by giving it a one or two line description of why you're talking to it.

4:14So I might say I am a 41 year old man, not really any past medical history. I've been undergoing a lot of stress lately and I have a pretty bad headache. Could you ask me some questions so that I can prepare to talk to my doctor about the headache? And that will have the model go through the way a doctor would ask you questions to collect that information. And then at the end of that interrogation, I might ask something like, can you prepare a short visit summary of the top three things I should tell my doctor and the top three questions that I should ask of my doctor? Really interesting. One of the warnings you give in the article is beware of sycophancy. In other words, you write the tendency of language models to try to please their users. Why is that a bad thing in a medical context? There's two reasons that it's a bad thing. So I think the first one is related to cyber-condria. And cyber-condria is a phenomenon that did not start with, it did not start with large

5:19language models. It's something that doctors and patients have been talking about and thinking about ever since, ever since the early days of the internet. And that's the tendency, right? If you search for a symptom, if you search for a diagnosis, the search algorithm is going to serve you up more and more concerning things. So very quickly, you can start searching for a headache and then soon be reading about, oh, could I have brain cancer? This same thing can happen with language models. Language models have all these safeties built into them. If you go to chat GPT these days and you ask at a health question, it is like 100% going to caveat. It's going to tell you talk to your doctor. It's going to have all of these warnings. But if you're worried about something, the language model is going to pick that up and the same phenomenon will happen. You'll start talking about a headache and very soon you'll be having a conversation about scary types of brain cancer, even though that's almost certainly what you don't have. So that's in that sense why sycophancy is bad. The second reason is, this is why I really like when my patients use it to prepare.

6:22This is one of the reasons that second opinions, even though they can theoretically be very good with language models, but sometimes they go wrong because if you're seeking out a second opinion and say you're putting your health information in there, there's a reason you're doing that. Maybe you're really anxious that you got a bad diagnosis and you don't want to have that diagnosis. These models can be very good at understanding what your concerns are and trying to please you. One of the fundamental tenets, of course, of the patient-doctor relationship is that it is a deep human relationship, and it's not just about making somebody happy. It's about doing what's best for them, and that nature of language models can really drive the wrong direction if you're not careful. Here's Julia in Hyatt's film, Maryland, who says she uses AI in a particular medical way. Hi, Julie, around WNYC. Hi, Brian, hi, Dr. Longtime Listener, several times caller.

7:23It's been invaluable for me in getting my A1C numbers down. What I usually do is I put my single-prick number in and also my sensor number, and it's also informed me about stacking, like protein first diet, and so it's been very helpful to me. Interesting. A1C, a reflection of the sugars in your system, diabetics, keep track of their A1C, or people concerned about their blood sugar going high. Drs, is that a familiar story to you? A hundred percent. Health coaching is one of the best things that these models are for. Looking at your A1C, reflecting on your diet, giving you advice, helping craft a personalized, personalized exercise program specific to you. These are the sorts of tasks that I would encourage my patients to think about. They're very, I would still want to know that my patients are doing this, but these are very low-harmed, high-potential benefit from having something that can counsel

8:27you at all times. Here's the text, listen to right. This sounds like a terrible privacy quagmire. Isn't it a bad idea to feed your personal medical data into the hands of AI companies? What do you say to that listener? The answer is yes. I feel conflicted about this, because these models are incredibly powerful for people understanding their health. They are also your personal health information. What I tell my patients, and this is what I do for myself if I have a question, is to take out any personal identifiers before you put it in. That's one of the reasons that I'm a little bit uneasy with what some of the tech companies are doing when they're automatically pulling from the health record. Obviously, you give permission, but they're automatically pulling that information in, which has identifiable information. If you do, and again, I'm not saying anyone should do this, but if you do do this, I would strongly encourage you to remove your personal identifiers. Now, the caveat to that is, in the LLM era, probably everything is potentially

9:27re-identifiable, so I understand where the listener is coming from, and I share that uneasiness. How are doctors using AI? That's another part of your article in the Times. Yeah. Doctors are using AI all the time. Survey data suggests that probably roughly two-thirds of doctors use AI on at least a weekly basis. There's really two main ways doctors are doing this. The first is what are called AI scribes, and many of your listeners have probably experienced this. You sit in the room, an AI listens to you talk to your doctor, and the doctor is not staring and typing at the computer, and then afterwards, the AI scribe writes the note. That's increasingly common, especially in primary care. The second is decision support. In the very old days, actually, when I was a medical student, we had these big books, physicians' desk reference. If we had a question that we go through, for most of my career, there have been websites that are curated by

10:27experts like up to date. Now, the way, actually, the most popular tool, and especially among younger doctors, among my residents, it probably has a 100% penetration. They all use it. Are these AI tools? The most popular one is called open evidence, but there's a couple. They're chatbots, and they're chatbots that have that search the literature for you. The odds that your doctor is using that is very, very high. It has gone from, I mean, it didn't exist a couple years ago, to being used by the majority of doctors in less than two years. Do you have warnings as an MD for your fellow MDs, as you have warnings for patients on how not to use AI? Oh, yes. The biggest warnings I have for my fellow MDs is that if you're getting complex information from a chatbot, you need to verify it, even though it's very good, and it has gotten much better, it's a language model. It hallucinates. It can say things that are not true or misrepresented paper, especially for very,

11:31very nuanced things. The bigger warning that I especially have for young doctors is that if you are using language, I mean, this is the problem, especially as an educator, language models are really powerful, and they actually can come up with very good recommendations. And part of medical training is to train your brain to be a doctor, to think like a doctor. And if you offload a lot of that at early stages of your career to an AI, you may never gain those abilities, and you may never be able to recognize when the AI is telling you something that doesn't make sense. One more call. Joe and Raway has a story of how one of his doctors uses AI, Joe and WNYC, hello. Hi, Brian, how are you? Yeah, I actually was on in a similar segment that was done one of the days that you were out. My primary care doctor actually uses an AI app called Medical Brain to communicate with his patients. And so it's a specialized AI app that he uses. And it keeps track or keeps in touch with me, how my health is, it checks my blood pressure,

12:36asks me for sugar numbers. And then one time I guess that the anecdote that I told before was it actually reminded me to check my blood sugars. I wasn't really good at that. I never really did unless I felt bad. And this was post-the procedure. And it came up really, really high. And I ended up contacting my doctors and we found some issues that were going on with that. So it seems to be really helpful. It's like a doctor checking in on you every day. And it also can communicate to the doctor if you send the message as well. And it's just beyond the one doctor now. It's going to a number of members of the team. So I wonder how that one is. As it said, it is pretty specialized. Yeah, Joe, thank you very much. Good story. We have 20 seconds left. Dr. Rodman, anything you want to say to Joe, but also do you recommend for people who are going to do this and take all your warnings and caveats about how and how not to use AI with respect to their health? Do you have any particular model company that you recommend?

13:40To answer your question first, no. Any of the companies with the Frontier model, so you know, your OpenAI models, your Google models, your Anthropics, those are all fantastic models. Everything has gotten so good. To Joe, I love that story. That is where I hope this feels feels goes to triadic care, where AI is an extension of the relationship you have with your doctor or your care team. And to me, that's my hope on where all this is going. Dr. Adam Rodman from Beth Israel Deaconess in Boston, Harvard Medical School, and his New York Times essay that had a few different titles on it, depending on where you were clicking around. One of them take it from a doctor. It's okay if your medical advice comes from AI, but really, it's a more complicated article than that with warnings as well as tips. Thank you for sharing it with us. Thank you, Brian. Thank you so much for having me. Brian Lera on WNYC, more to come. Since WNYC's first broadcast in 1924, we've been dedicated to creating the kind of content we

14:45know the world needs. Since then, New York Public Radio's rigorous journalism has gone on to win a Peabody Award and a Dupont Columbia Award, among others. In addition to this award-winning reporting, your sponsorship also supports inspiring storytelling and extraordinary music that is free and accessible to all. To get in touch and find out more, visit sponsorship.wnyc.org.