#255 Eric Topol: Why AI is the Most Powerful Tool in Healthcare Now

0:00Now things are different. First of all you have all these different layers that are orthogonal that back up each one from another, looking to the same, try that forecast. I call it precision medical forecast. Like the weather forecasting, AI is completely revamping that, making it at accuracy levels that are hard to fathom how great it will be. The same thing is true here. So basically the difference is you've got all these additional layers of data we didn't have. And what's really noteworthy is you can give a person's projection in time when this is going to be an issue, if they don't do anything about it.

0:32about the science of aging. It's not about the drug. It's about being able to clock a person, a person and their organ systems, including their immune system. We never had that before. That's what's different right now. AI is everywhere right now, but with all the buzz, how do you cut through the noise and focus on what really matters? If you're looking to dive deep into the big questions shaping AI's role in business today, you need to tune in to where AI works. Conversations at the intersection of AI and industry. Brought to you by the Wharton School in collaboration with Accenture. Each episode cuts through the hype, blending cutting-edge research from Wharton professors with real-world case studies.

1:12You will discover how leading companies are leveraging AI to upskill their teams, boost productivity, and streamline operations. And you will hear directly from executives and industry pioneers who are implementing AI in their businesses today. No fluff, just practical insights. Listen to Where AI Works Now on your favorite podcast app and get the strategies that matter most for today's business leaders. Introduce yourself to listeners. Obviously, you're a doctor, but how you got so deeply involved in the tech space and what your book is about, and then we'll talk about that. Well, I'm Dr. Eric Topol.

1:54I am a professor and EVP at Scripps Research in La Jolla. California. I've been involved from the earliest days of digital medicine back in the 90s with biosensors, with smartphones connected to the internet, even before there was a real digital infrastructure. So back many years ago, I wrote a book called The Creative Destruction of Medicine, which was about going digital in medicine. And then most recently about to be published as a book called Super Agers. It's actually about using AI to promote healthspan. It has not been previously laid out like that. It's a blueprint for how we can go forward.

2:42And we couldn't do it without the current types of models with AI. And that's what makes it especially exciting. Yeah, I just saw Bill Gates at an interview talking about how someday soon within the next decade, we're all going to be accessing medical information through agents, AI agents. uh so super agers the the uh yeah give me the thesis of the book is it is it about how ai is going to spread uh knowledge and and uh or is it more about uh aging strategies or how to combat aging yeah so there are the three major uh age-related diseases uh heart um neurodegenerative alzheimer's and cancer they all take 20 years to incubate they all have common threads of the immune system and inflammation and now we can tell in any given individual who is at high risk by a full stack of data depth that we've never had before that's not including not just including electronic health records and labs and images but it's also including the genome proteins metabolites they get microbiome environmental exposures social determinants of health anyway for all these different layers of data the ai will tell us if you are high risk when Then pinpointing when this will show up if nothing is done.

4:30And then that can be used to direct surveillance and prevention strategies. So you never have one of those three age-related diseases, which are the biggest burden of our chronic disease and killers, of course. So this is an exciting time. It's unique. it's it's momentous in many respects because we didn't have the ai that could could input ingest all this data and we didn't have the data to ingest right so um what we have now is a is a just extraordinary time to be able to help each person know especially at a younger age what's in store for them and to do these things that would help prevent at the very least markedly delay a condition yeah and that data i mean there are two issues one is uh the the privacy issue that's kept a lot of uh health care data locked up and inaccessible to ai models but also personal health data that is not necessarily digitized or or very complete longitudinal i I mean, I've moved around the country.

5:53I don't know where all my, not the country, the world, where all my data is, if it exists. How do you tackle those two problems? I mean, first of all, on just more broadly, in order to train models, you need access to large populations, data of large populations. And I've spoken to people over the years about that. It seems that that's been addressed to a certain extent. But, yeah, where do you see that today and how it's going to improve? Okay, well, firstly, the data story that you bring up is that I've written about extensively, even had a book about it, The Patient Will See Now. people should have all their data it's their data they should own it and it should be kept in a secure platform and so other countries do that just not here in the US but that's the ideal scenario because you're not going to put your genome sequencing data or your biosensor data or other layers of data into your electronic health record because we don't protect people fully for things like disability insurance and life insurance and you know various things that are not related to health insurance and the so-called gina you know discrimination of people about their genetic information so we have to come up with a better plan about where all that data sits and that each person is entitled it's their data their body their that's critical now it turns out the am models that we have now um are very well suited for um supervised fine tuning for this task i mean the fact is they're multimodal uh they're now foundation um we can use of course uh domain specific agents, but no problem here with respect to the models.

8:07If you were talking pre-transformer architecture, yes, but in the current environment, we even are getting these reasoning capabilities. And it's not like a lot of these things have to be outputs in seconds. We can wait minutes, hours, days, even for the best output. So the models are not the holdup. The data ownership is actually an issue that has to be grappled with. In the meantime, though, many people have been able to collect all their data, curate everything of theirs, particularly engineers and certain people, you know, patients of mine that have everything. And so they're very well suited to enact this now.

8:55and um is okay so that's the personal data what about this this more universal data to train the models yeah it's surprising that uh it's not as big a challenge as you would expect because the it a lot of the things in medicine and life science uh we would never predicted that the models that already were appearing like in gbt4 and and llama and others really were remarkably well suited for for this uh just by having ingested everything known to mankind about medicine and um um you know the whole uh corpus of knowledge related to this So that, you know, in the work that's been done so far to look at being able to predict the arc of a person's, let's say, risk for Alzheimer's and when they would get out, we can do that now.

9:58We don't need new models, in fact. So the difference now is, Craig, is that up until now, let's say you had a so-called polygenic risk score where you could go to 23ME or a whole bunch of different companies and tell you, you know, Craig, you have a risk of, it's high for, let's say, Alzheimer's disease. And the only problem with that is, well, there are two problems. One is it might not be that accurate. But the bigger problem is, what if the risk is when you're 98 rather than 68? Big difference. And there was no way to be able to ascribe a time range. Now things are different. First of all, you have all these different layers that are orthogonal, that back up each one from another, looking to the same, try to forecast.

10:49I call precision medical forecasting. Like the weather forecasting, AI is completely revamping that, making it at accuracy levels that are hard to fathom how great it will be. The same thing is true here. So basically the difference is you've got all these additional layers of data we didn't have, and what's really noteworthy is you can give a person's projection in time when this is going to be an issue if they don't do anything about it. Now, many of the interventions or preventive tactics involve lifestyle changes. And we know a lot of those, but most people don't do any of those or change behaviors.

11:33But we've also learned that when you have specific knowledge about yourself, the chances of you taking on some of these changes are much higher. So these are now much more specifically bespoke to your situation. But by the way, it isn't just lifestyle changes that can help prevent these conditions. I mean, obviously there's surveillance, whether it's cancer surveillance with multi-cancer early detection blood tests, or whether it's surveillance for neurodegeneration with these new p-tower 217 biomarkers, which was a really big breakthrough. Each of these disease, age-related diseases, has ways we can stay ahead of it.

12:19Yeah. It just reminds me, as an aside, so this is really about prediction. Yes. I've been talking over the years to Alex Jovronikov from InSilico. I'm sure you know of him if you don't know. Oh yeah. No, I know him. He sends me stuff all the time. Yeah. Well, and his, you know, he's all on, uh, you know, anti-aging, uh, and that's kind of his motivation. And he's, you know, he's got a very sophisticated, uh, process for drug discovery and he does. Yeah. Uh, and there he's working on inflammatory diseases because, uh, he sees that as one of the major manifestations of aging that end up leading to death.

13:18Yeah, fibrosis. Yeah. Have you looked at, in the book, do you look at using this technology to develop drugs to target specific diseases uh of uh that are endemic to aging yeah so this is um i do review all the different strategies uh that are now being pursued um and you know everything from uh partial epigenetic reprogramming to you name or you know different repair um like what in silico medicine is on to with pulmonary fibrosis. You know, I review all that, but basically the conclusion is these are years away, if ever. You know, things like senolytics to take out our senescent cells in our body, but they can't tell the good senescent cells from the bad senescent cells.

14:15And a lot of these things, you know, induce cancer in some organisms. So they're risky, They're iffy, they're off in time, but today what we can achieve is using AI tools and rich data, deep data for each person to forecast like we've never been able to do before. And so one of the biggest breakthroughs were these so-called organ clocks. So where Alex and I and many others in the anti-aging field converge is there have been dramatic advances in the science of aging. And what is an outgrowth of that are these things, these molecular clocks, like epigenetic clocks, like the Horovath clock, where you look at methyl groups on the DNA.

15:03And there's also organ clocks. From a tube of blood, you can get 11 ,000 plasma proteins. It'll tell you which organ. So I'd say, Craig, all your organs are great, but it turns out your heart is, you know, seven years accelerated aging, which tells us right away, you know, that's what we got to zoom in and for you. So we're seeing tools that we never had either breakthrough tests about the science of aging. It's not about the drug. It's about being able to, you know, clock a person, a person and their organ systems, including their immune system. We never had that before. That's what's different right now.

15:45And with these different proteins and assays, we can say, hmm, we measured it this year, we measured it, you know, this year, all that means for you to be worried about this, it's if you if you don't really go full court press, at age, you know, whatever 71, this is going to hit you, you're going to start having symptoms. So with that point that we have 20 years to undo the natural history of risk, it's a it's an, you know, this is a whole different strategy. The book is about a different strategy than the one that these companies are pursuing, like Altos Labs and Unity and in silico, you know, there's 50 of these anti-aging companies, right?

16:31I'm saying And good, good luck to them. I hope they're successful. But I have a different plan that's not risky, okay? Because preventing diseases is a dream in medicine that's been a fantasy that's been around for, you know, multiple millennia. Now we have a chance to do something about it. Yeah. How much of this is productized at this point? Everything's like, yeah, you're right. Everything's piecemeal. So you got to go here to get your organ clock. You got to go here to get your, your epigenetic clock. You got to go here to get your metabolome, you know, here to go to get your gut microbiome.

17:16So it isn't, the package isn't, but you could do it. You can do it. It's just that the, the AI tool is separated from each of the layers of data. It's just a matter of time when this comes together. Your point is a good one, Craig. Yeah. I just wrote a piece for Forbes about MCP. I'm not going to remember what it stands for, but this Anthropic developed standard protocol for connecting AI systems to disparate data sources or disparate tools sort of thing. This one is being cracked every day with multi-omic. So in the life science field, which is what this is really converging with, they are already looking at every layer of DNA, RNA, single sequence, single cell sequence, nucleus, and all the other ones I've mentioned.

18:15They've already figured out how to do all that. And that's exciting because that's just part of this. all you got to do is you know add in a couple of traditional medical pieces like the electronic records which contain labs and images and you're there yeah uh in the book do you go through how somebody could collect all of this data because for the lay person i would i mean just the the list that you mentioned uh i don't know anything about did you say melanome Oh, the metabolome. Metabolome. Yeah. I mean, I don't know what that is, where it is, how I could find it. Yeah. Well, there's a company, Metabolon, that does that and other companies as well.

19:05I don't go through that step by step. I just basically trying to say, look, folks, we're going to a new era where you're going to know your risk. And I lay out, you know, what are the layers of data, how we're going to do this. In the next year or two, we'll see the companies that have the whole package. Right now, there's these companies pitching data for longevity. And I called them out. I wrote a sub stack about that recently. You know, these companies that are selling supplements. I mean, this is not my science. But they have the right idea, okay? Because some of them are assembling these different layers of data already.

19:49They're not using AI. Sometimes they say on their website they use AI, but no, they're not real AI. So we're already seeing the budding. You know, there's several companies that this is in their sights right now, and it's not going to take long for them to adapt to the opportunity, which, you know, this is a very clear path to achieving something we've never been able to do before. Our ability to give a person's risk assessment of a disease has been rudimentary at best, but actually incredibly weak and inaccurate. And it's our most important thing is that we should be knowing well in advance what we have to prevent and how we do it.

20:38Yeah. And one thing you said, I was thinking as you were talking, you know, I've looked a lot at ed tech and I've spoken to a couple of companies that have an assembly of algorithms and one of the things that struck me as most promising is that you can have a knowledge tracing algorithm that tracks the knowledge of a student as the student studies provided that they're interacting digitally. And then you can have a prediction algorithm that predicts the student's score in an exam or in a course. And then you can have a recommendation engine that recommends stuff to fill in the knowledge gaps. And when you put it all together, I mean, theoretically, the student, if he follows the recommendations, can see his predicted score going up, and that's a huge motivator.

21:43It could be the same thing with... Same thing. Same exact model type of... I agree. More compelling because it's your lifespan that you're increasing, not... Yeah, exactly. And just to add to it, so you know the likes of Jeff Hinton, Demet Pesabas, Mustafa Suleiman. All three of them reviewed the book. And they all thought this is, you know, extraordinary. They supplied, you know, endorsements, not, you know, because they see that, you know, actually both Demis and Jeff have said that the most important ever use of AI will be to improve our health. And this is, I think, unless I don't know, Craig, but I don't think anybody has laid out how we're going to do this yet.

22:33This is kind of that's why I'm excited about it. and they were excited about it. So, you know, we have now a way, it's very similar to what you just described in education. Of course, it's a totally different application. It's a little more complex, but it's ready to go. It's not going to take long to get there. Building multi-agent software is hard. Agent-to-agent and agent-to-tool communication is still the Wild West. How do you achieve accuracy and consistency in non-deterministic agentic apps? That's where agency comes in. A-G-N-T-C-Y. The agency is an open source collective building the Internet of Agents.

23:24And what's the Internet of Agents? It's a collaboration layer where AI agents can communicate, discover each other, and work across frameworks. For developers, this means standardized agent discovery tools, seamless protocols for interagent communication, and modular components to compose and scale multi-agent workflows. Build with other engineers who care about high-quality multi-agent software. Build with other engineers who care about high-quality multi-agent software. Visit agency.org and add your support. That's A-G-N-T-C-Y dot O-R-G. Yeah. Are you free to mention companies that you think are on the right track in coming out with an application?

24:21Well, I reviewed all of them that are out there. um and you know there i'll just pull up the chart i made um from the ground truth that i did uh one sec here if you can bear with me um sure uh yeah i just put it out a week ago i called it the business business of promoting longevity and healthspan basically the 12 companies there was one that started this years ago a decade ago called human longevity they get total body mri and genome sequence this was before there was any ai i got all these labs and whatnot that company is just basically not almost non-existent now charging 25 000 to come and have all these things done that's not that's a no-brainer not going to work uh function health they do 100 lab tests that's it you know so that's not exactly what we're talking about um and of the ones life force is um i think on the right track um let's see going down the list um fountain life is trying to get a whole lot of tests actually when you look at them they have the right idea that you need more layers of data, but none of them are getting the right data.

25:45I wrote about that in the post. They have this idea that, oh, if I know much more data about a person, I can guide them. They're just not getting the right data. And none of these companies have a significant AI effort yet. But some of them have raised immense capital. Like, for example, function has has a valuation of two and a half billion dollars already it has uh 50 000 members and a wait list uh neco health has a hundred thousand people on a wait list so people want this yeah okay uh but they they're not they're a little clueless about they're not getting what they should get yeah are you advising i mean you obviously are deep in this not advising any of them.

26:32I don't want to be conflicted. Well, why you have a deep knowledge of AI and you have a deep knowledge of medicine. Have you thought about doing exactly what we're talking about? Yeah. Okay. So the function health, the one that has this massive valuation, one of the principles is a doctor named Mark Hyman. He flew here, right? He was sitting in my office and he offered me to, you know, be part of the company and, you know, get all sorts of financial incentives. I told them, Mark, I just can't do that. Because if I do that, then people will read this book, or see me as, you know, I'm doing this for financial gain, just like they sell supplement, which I don't agree with.

27:19That's another reason I wouldn't do it. Right? So basically, I can't, for me, it's much more important to put out the ideas and not have any profit motive whatsoever. whatever yeah yeah yeah well i didn't mean as much as as profit motive i mean obviously you're you've had a financially lucrative career uh but just as you you see the problem you see the solution you have the knowledge uh and certainly the contacts to pull together a team that that could solve the problem well i i just wanted to get it done and like i have done in the other previous books i gave a lot of people the ideas right i so many people have written me or come visit me and tell me you know i built my company because of what you wrote and that's kind of what i'm hoping is other people can execute here's the here's the recipe here's the blueprint whatever you want to call it and that's i think how it's going to get done now the reason i mentioned these companies they already have a head start they have revenue they have you know some some of them have pretty strong participation memberships whatever so i i i'm perfectly happy to see them pivot do this right get there where we eventually all will be i don't want this to be a thing just for the affluent either yeah in fact the people who need this the most are the ones who can't access or afford it yeah uh how much of this i mean you you understand the relevant data how to get the relevant data how much of that have you done yourself uh most of it so in the book i got uh an immunom the beginning of an immunom from a company infinity bio john's hopkins spinoff which is another essential part of the layers of data uh i've had most of the other layers of data you know, so I have done this.

29:19And the only risk of this big three that I'm facing is the cardiovascular risk. And so now I know what I got to do, I'm going to get my LDL down to the lowest level. And, you know, 10 other things that I got to do. But there are things that, you know, turning to much better sleep health, getting much better deep sleep, and many other things that we can do once we know what we're after. Now, if you find that you don't have a risk for one of these big three diseases, well, you're lucky, right? Doesn't mean you should go out and go into self-destruct mode, but you're lucky. The point is most people, one of them, if not more than one, have vulnerability.

30:03And the question is the cutoff. Top 5%, top 10 that's something that you know is still going to be debatable for years to come what is a risk right well and and you were talking about the timing predicting the timing of risk uh how accurate do you think that can be i mean yeah i think it can be it's hard to get it on one shot um but if you have a person that you see is at high risk by the time you do a second assessment then you have it so if you have just two points on the curve uh you'll know what's about the brain amyloid accumulation what about microscopic cancer showing up what about atheroma forming in arteries so yeah it doesn't take it's hard to do it on a single uh point basis but two data points separated in time by six months or a year and you nailed it the key is that first baseline to determine the risk and then use the second point to determine the timing.

31:10Yeah. And for the timing, for example, let's pick one, cancer. Yeah, yeah. You know, you look at your family history. You look at presumably different biomarkers, maybe your personal genome. um what would what would the prediction be that that in three years in 15 20 years you you have an uh you know a 95 chance i mean what does the prediction look like okay well the 95 percent that's what we were talking about earlier top five percent that would be 95 percent or very high risk that is um you know somewhere i'm thinking you know most people if they're not in the top five or 10 percent they're not going to really it's that's really where you draw the threshold that's where some medical health systems like for example mass general brigham has drawn the threshold for genetic risk so let's just say we're talking about top five or ten percent we're not likely to get out a readout that you know this is going to happen in 18 years but certainly you know when it's more proximal like three to five years that would be the way show up.

32:27Now, how do we get this? Well, it turns out, and I make a really important point early in the book, we have overestimated the family history genetic story. Okay. Now for cancer, it's more of an issue than it is for the other two, because if you have any of these cancer predisposition genes, we're not just talking about BRCA genes, but other genes, that puts you in a different strata of risk. But the point being here is that there are many other factors like the level of your body inflammation, that those organ clocks, again, your total body aging clock. If you're seeing things that are out of kilter, and then you get a multi-cancer early detection test, the chance of that being abnormal.

33:19See, today, what's amazing, these tests are advocated for people age 50 period now talk about a senseless way to use a test you know out of a thousand people that get the test now 995 are negative yeah okay but when you use that test when you have these other things that are um suggesting increased risk then the yield is so much higher. I mean, this is Bayes' theorem. So that's how we get the risk of cancer nailed down. And then, of course, if there's an organ clock that's already showing up, you can start to, and you're going to get images, of course, like a total body MRI, if you really, but the multicenter cancer detection test already with AI can tell you which organ is lighting up most of the time very accurate actually so the cancer story i think is not that hard to see through um as long as you don't say it's just the genes that's our biggest problem is we think we we have this kind of over uh emphasis about our family history and this kind of dooms uh impending doom thing that we live our whole life when my father had a heart attack when is 50 and you think about that like all the time we got to get that out of our system because that just doesn't really play out we did a so-called welderly study we called it here 1400 people lived through past age 85 never sick none of these diseases we sequenced whole genome all of them and what did we find?

35:00Very little, almost nothing. So the secret in the genes thing is not so, there's not that much to it. There's a little, but not much. Yeah.

35:16The, let me think, where do we go with this? the conclusion of the book then is that there's this promising technology that's here and there's ways to get the data to make the technology useful. So what are you, this is, I would presume, not only for AI technicians, but for general readers. What is the recommendation? I mean, what is the book telling us to do? Well, there's a whole chapter on these lifestyle things that, you know, it'd be great if we adopted all those now. Right. Because, you know, without even knowing your risk for a particular condition, it'd be good. But I review all the data, you know, everything from, you know, protein and ultra processed foods and environmental hazards, microplastics.

36:25I mean, all this stuff. Each of us have at least some potential to affect. Right. And then I get into this, what I call, you know, brimming with optimism, that we have a much more, you know, exciting way to go forward in medicine, to be able to be informative early in a person's life. And I mean, it's never too late, but for people of our age, but even better is, you know, somebody in their 40s or 50s or younger even. But the point here is that there's a new path forward that I think is transformative in medicine. We couldn't do it without multimodal AI models that exist today and keep getting better.

37:12And we couldn't do it without the layers of data that were discovered largely through the science of aging. But a lot of people are using that, you know, of course, for companies to come up with drugs. And I'm saying, no, no, no. We got it for a different purpose now. And let's stamp out age-related diseases. Let's do it. So the book leaves the lay reader, I hope, with an optimistic view in otherwise pretty dark times for life science right now when, you know, our funding is getting gutted every day. People are getting, you know, laid off from NIH, FDA, CDC. So on the one hand, the world isn't doing too well right now.

37:51But here's something that's hopefully, you know, exciting to read about, know where we're headed, I mean, each of the books I've written, this is my fourth book, probably my last book, is about the future of medicine. And this one is the most exciting yet. This one, I think, is something we can really make a huge difference in the future. Yeah. On, you know, there's a lot of talk about personalized medicine and with with you know the ability now to to edit the genome and create novel proteins uh what's your view just generally i know this is not what your book's about, but about the promise of precision personalized medicine becoming widely available.