Ground Truths

*/ ?>

Latest Episodes:

John Halamka: How Mayo Clinic is Transforming Healthcare with A.I. Aug 11, 2023

Show notes

Transcript Eric Topol (00:00):This is a real great opportunity to speak to one of the most impressive medical informaticists and leaders in AI in the United States and worldwide. Dr. John Halamka, just by way of background, John, his baccalaureate in Stanford was at U C S F/Berkeley for combined MD PhD trained in emergency medicine at U C L A. He went on to Harvard where he, for 20 years was the Chief Information Officer at Beth Israel Deaconess. And then in 2020 he joined Mayo Clinic to head its platform to help transform Mayo Clinic to be the global leader in digital healthcare. So welcome, John. It's so great to have you. And by the way, I want to mention your recent book came out in April, one of many books you've written, redefining the Boundaries of Medicine, the High Tech High Touch Path into the Future.John Halamka (01:00):Well, a thrilled to be with you today, and you and I need to spend more time together very clearly.Eric Topol (01:06):Yeah, I really think so. Because this is the first time we've had a one-on-one conversation. We've been on panels together, but that's not enough. We've got to really do some brainstorming, the two of us. But first I wanted to get into, because you have been on a leading edge of ai and Mayo is doing big things in this space, what are you excited about? Where do you think things are right now?John Halamka (01:35):So you and I have been in academic healthcare for decades, and we know there's some brilliant people, well-meaning people, but sometimes the agility to innovate isn't quite there, whether it's a fear of failure, it's the process of getting things approved. So the question of course is can you build to scale the technology and the processes and change policies so that anyone can do what they want much more rapidly? And so what's been exciting over these last couple of years at Mayo is we started with the data and we know that anything we do, whether it's predictive or regenerative, starts with high quality curated data. And so by de-identifying all the multimodal data of Mayo and then working with other partners around the world to create a distributed federated approach for anyone to train anything, suddenly you're empowering a very large number of innovators. And then you've seen what's happened in society. I mean, culturally, people are starting to say, wow, this ai, it could actually reduce burden, it could democratize access to knowledge. I actually think that yes, there need to be guidelines and guardrails, but on the whole, this could be very good. So here we have a perfect storm, the technology, the policy, the cultural change, and therefore these next couple of years are going to be really productive.Implementing a Mayo Randomized AI TrialEric Topol (02:59):Well, and especially at Mayo, the reason I say that is not only do they recruit you, having had a couple of decades of experience in a Harvard program, but Mayo's depth of patient care is extraordinary. And so that gets me to, for example, you did a randomized trial at Mayo Clinic, which there aren't that many of by the way in AI where you gave E C G reading power of AI to half the primary care doctors and the other half you didn't for determining whether the patients had poor cardiac function that is low ejection fraction. And now as I understand it, having done that randomized trial published it, you've implemented that throughout the Mayo Clinic system as far as this AI ECG support. Is that true?John Halamka (03:56):Well, right, and let me just give you a personal example that shows you how it's used. So I have an SVT [supraventricular tachycardia] , and that means at times my resting heart rate of 55 goes to one 70. It's uncomfortable. It's not life-threatening. I was really concerned, oh, may I have underlying cardiomyopathy, valvular disease, coronary artery disease. So Paul Friedman and Peter Newsworthy said, Hey, we're going to take a six lead ECG wearable, send it to your home and just record a bunch of data and your activities of daily living. And then we buy 5G cell phone. We'll be collecting those six leads and we'll run it through all of our various validated AI systems. And then we'll tell you based on what the AI suggests, whether you're at high risk or not for various disease states. So it says your ejection fraction 70%. Oh, good. Don't have to worry about that. Your likelihood of developing AFib 3% cardiomyopathy, 2% valvular disease, 1%. So bottom line is without even going to a bricks and mortar facility here, I have these validated algorithms, at least doing a screen to see where maybe I should get additional evaluation and not.Eric Topol (05:12):Yeah, well see what you're bringing up is a whole other dimension. So on the one hand that what we talked about was you could give the primary care doctors who don't read electrocardiograms very well, you give them supercharged by having a deep learning interpretation set for them. But on the other, now you're bringing up this other patient facing story where you're taking a cardiogram when somebody's perfectly fine. But from that, from having deep learning of cardiograms, millions of cardiograms, you're telling what their risks are that they could develop things like atrial fibrillation. So this is starting to span the gamut of what the phase that we went through or still going through, which is taking medical images, whether it's a cardiogram or a scan of some sort, and seeing things with machines that humanize really can't detect or perceive. So yeah, we're just starting to get out of the block here, John. And you've already brought up a couple of major applications that we were not even potentially used three, four or five years ago that Mayo Clinics leading the charge, right?The Power of Machine EyesJohn Halamka (06:26):Well, yeah, and let me just give you two quick other examples of these are in studies now, right? So they're not ready for active patient use. The animate GI product does an overread of endoscopy. And what we're finding is that the expert human, I mean anywhere in the world, expert humans miss about 15% of small polyps. They're just hard to see. Prep may not be perfect, et cetera. The machine misses about 3%. So that's to say a human augmented with overread is five times better than a human alone pancreatic cancer, my father-in-law died about 11 years ago of stage four pancreatic cancer. So this is something that I'm very sensitive about, very often diagnosed late, and you can't do much. What we've been able to see is looking at pancreatic cancer, early films that were taken, abdominal CT scans and these sorts of things, algorithms can detect pancreatic cancer two years before it is manifested clinically. And so here's the ethical question I'll pose to you. I know you think about a lot of this Scripps Mayo, UCSF, Stanford, we probably have thousands and thousands of abdominal CTs that were read normal. Is it an ethical imperative as these things go through clinical trials and are validated and FDA approved to rerun algorithms on previous patients to diagnose disease we didn't see?Eric Topol (08:03):Well, that is a really big important question because basically we're relieving all this stuff on the table that doesn't get diagnosed, can't be predicted because we're not even looking for it. And now whether it's retina, that is a gateway to so many systems of the body, or as you're mentioning various scans like an abdominal CT and many others that like mammography for heart disease risk and all sorts of things that weren't even contemplated that machine eyes can do. So it's really pretty striking and upending cancer diagnosis, being able to understand the risk of any individual for particular types of cancer so that you can catch it at the earliest possible time when it's microscopic before it spreads. This, of course, is a cardinal objective. People don't die of cancer per se. They die of its metastasis, of course, for the most part. So that gets me now to the next phase of ai because what we've been talking for mostly so far has been what has been brewing culminating for the last five years, which is medical images and what, there's so many things we can glean from them that humans can't including expert humans in whatever discipline of medicine.Multimodal AI and Social Determinants of Health(09:19):But the next phase, which you are starting to get at is the multimodal phase where you're not just taking the images, you're taking the medical records, the EHRs, you're getting the genomics, the gut microbiome, the sensors. You mentioned one, an ECGs, a cardiogram sensor, but other sensors like on the wrist, you're getting the environmental things like air pollution, air quality and various things. You're getting the whole ball of wax any given individual. Now, that's kind of where we're headed. Are you doing multimodal ai? Have you already embarked in that new path? Now that we have these large language modelsJohn Halamka (10:02):And we have, and so like anything we do in healthcare innovation, you need a Pareto diagram to say, what do you start with and where do you go? So in 2020, we started with all of the structured data problems, meds, allergies, labs. Then we went to the unstructured data, billions of notes, op reports, H and Ps, and then we moved to telemetry, and then we moved to CT, MRI, PET. Then we move to radiation oncology and looking at all the auto contouring profiles used in linear accelerators and then to omic, and now we're moving to an inferred social determinants of health. And let me explain that for a minute.(10:45):Exposome, as you point out, is really critical. Now, do you know if you live in a Superfund site area, do you know what risks you might have from the PM 2.5 particulates that are blowing through San Diego? Probably you don't. So you're not going to self-report this stuff. And so we have created something called the house Index where we've taken every address in the United States, and based on the latitude and longitude of where you live, we have mapped air, water, land, pollution, access to primary care, crime, education, grocery stores, stores, and therefore we can infer about 40 different things about your expose em just from where you live. And that's a mode. And then as you say, now, starting to gather remote patient monitoring. We have this acute advanced care in the home program where we're taking serious and complex illness, caring for the patient in the home, starting to instrument homes and gather a lot more telemetry. All of that multimodal data is now available to any one of the 76,000 employees of Mayo and our partners for use in algorithm development.Eric Topol (11:58):Yeah, no, that's extraordinary. And I also would say the social determinants of health, which you've really gotten into as its importance. There are so many papers now over the last several years that have emphasized that your zip code is one of the most important things of your health. And it's not even just a zip code. It's your neighborhood within that zip code for the reasons that you've mentioned. And inferring that and imputing that with other sources of data is vital. Now, this multimodal, you've again anticipated one of my questions, the possibility that we can gut hospitals as we know them today. Yes, preserving the ICUs, the emergency departments, the operating rooms, but those other people that occupy the vast majority of beds in the hospital that are not very sick, critically Ill. Do you think we're going to move as you're innovating at Mayo whereby we'll be able to keep those people at home for the most part in the years ahead? I mean, this isn't going to happen overnight, but do you think that's where we're headed?The Hospital-at-HomeJohn Halamka (13:08):So to date, Mayo and its partners have discharged about 23,000 patients from their homes. And as you can guess, we have done clinical trials and deep dive studies on every one of the patient's journeys. And what have we seen across 23,000 patients? Well, so generally, about 30% of patients that present for acute care to an emergency department come in by ambulance are appropriate for care in non-traditional settings. I mean, I think you would agree, somebody with episodic ventricular tachycardia, you're probably not going to put in a home setting, but somebody with congestive heart failure, COPD, pneumonia, I mean, these are things that, as you say, if they're going to get sicker, it will be over hours, not minutes. And therefore you can adjust in these molar than 20,000 patients. What we've seen is the outcomes are the same, the quality is the same safety, the same patient satisfaction. You get net promoter scores in the mid-nineties. You find me a hospital with a net promoter score in the mid nineties. You're eating your own food, slipping your own bed. Oh, your granddaughter's coming at 2:00 AM on a Sunday, whatever. And then ask yourself this other question, nosocomial infections,Eric Topol (14:31):Right?John Halamka (14:31):How many methicillin resistant staph infections do you have in your office? You're like, none, right? So you're infections in fall, so okay, better, stronger, cheaper, faster. And the safety of the quality are that for about 30% of the population should be a standard of care.Eric Topol (14:56):That's really big. So you don't think we have to do randomized trials to prove it?John Halamka (15:01):I mean, we have done enough studies to date, and there are organizations, Kaiser Permanente, Cleveland Clinic, all these folks who are joining us in investigating these areas. And the data is very compelling.Patients Asking Questions to LLMsEric Topol (15:17):Yeah, that's really exciting. And we may be able to jump past having to go through the large trials to prove what you just reviewed. So that's one thing of course that we're looking for in store. Another is the patient doing advanced large language model searches. So as you and everyone knows, we've done Google searches for years about symptoms, and inevitably people come up with hypochondria because they have some horrible disease that they looked up that is not a very good match specific for their condition and their background. But soon already today, we have people going into being creative mode, G P T four and other searches, and they're getting searches about their diagnosis and about what's the best literature and best treatments and expectations. That won't be FDA regulated. We don't have regulation of Google searches. So how do you see the democratization of large language models with patients having conversations with these chatbots?John Halamka (16:32):And of course, you ask a question no one has answered yet, but here are a few threads. So we know the challenge with existent commercial models as they're trained on the public internet. Some are trained on additional literature like PubMed or a mimic dataset, but none are trained on the rich clinical experience of millions and millions of patients. So therefore, they don't have the mastery of the care journey. So question, we are all asking, and again, no one knows. Then you take a GPT, BARD, a MedPaLM and additional pre-training with rich de-identified clinical experience and make it a better model for patients who are going to ask questions. We've got to try and we've got to try within guardrails and guidelines, but we definitely want to explore that. Can you or sho…

Full show notes at the publisher

Melanie Mitchell: Straight Talk on A.I. Large Language Models Aug 04, 2023

Show notes

Transcript with LinksEric Topol (00:00):This is Eric Topol, and I'm so excited to have the chance to speak to Melanie Mitchell. Melanie is the Davis Professor of Complexity at the Santa Fe Institute in New Mexico. And I look to her as one of the real, not just leaders, but one with balance and thoughtfulness in the high velocity AI world of large language models that we live in. And just by way of introduction, the way I got to first meet Professor Mitchell was through her book, Artificial Intelligence, A Guide for Thinking Humans. And it sure got me thinking back about four years ago. So welcome, Melanie.Melanie Mitchell (00:41):Thanks Eric. It's great to be here.The Lead Up to ChatGPT via Transformer ModelsEric Topol (00:43):Yeah. There's so much to talk about and you've been right in the middle of many of these things, so that's what makes it especially fun. I thought we'd start off a little bit of history, because when we both were writing books about AI back in 2019 publishing the world kind of changed since . And in November when ChatGPT got out there, it signaled there was this big thing called transformer model. And I don't think many people really know the difference between a transformer model, which had been around for a while, but maybe hadn't come to the surface versus what were just the deep neural networks that ushered in deep learning that you had so systematically addressed in your book.Melanie Mitchell (01:29):Right. Yeah. Transformers are, were kind of a new thing. I can't remember exactly when they came out, maybe 2018, something like that, right from Google. They were an architecture that showed that you didn't really need to have a recurrent neural network in order to deal with language. So that was one of the earlier things, you know, and Google translate and other language processing systems, people were using recurrent neural networks, networks that sort of had feedback from one time step to the next. But now we have the transformers, which instead use what they call an attention mechanism where the entire text that the system is dealing with is available all at once. And the name of the paper, in fact was Attention is All You need. And that by attention is all you need they meant this particular attention mechanism in the neural network, and that was really a revolution and enabled this new era of large language models.Eric Topol (02:34):Yeah. And as you aptly pointed out, that was in, that was five years ago. And then it took like, oh, five years for it to become in the public domain of Chat GPT. So what was going on in the background?Melanie Mitchell (02:49):Well, you know, the idea of language models (LLMs) that is neural network language models that learn by trying to predict the next word in a, in a text had been around for a long time. You know, we now have GPT-4, which is what's underlying at least some of ChatGPT, but there was GPT-1 and GPT-2, you probably remember that. And all of this was going on over those many years. And I think that those of us in the field have seen more of a progression with the increase in abilities of these increasingly large, large language models. that has really been an evolution. But I think the general public didn't have access to them and ChatGPT was the first one that like, was generally available, and that's why it sort of seemed to appear out of nothing.SPARKS OF ARTIFICIAL GENERAL INTELLIGENCESentience vs IntelligenceEric Topol (03:50):Alright. So it was kind of the, the inside world of the computer science kinda saw a more natural progression, but people were not knowing that LLMs were on the move. They were kinda stunned that, oh, look at these conversations I can have and how, how humanoid it seemed. Yeah. And you'll recall there was a fairly well-publicized event where a Google employee back I think last fall was, put on suspension, ultimately left Google because he felt that the AI was sentient. Maybe you'd want to comment that because that's kind of a precursor to some of the other things we're going to discuss,Melanie Mitchell (04:35):Right? So yeah, so one of the engineers who was working with their version of ChatGPT, which I think at the time was called LaMDA was having conversations with it and came to the conclusion that it was sentient, whatever that means, , you know, that, that it was aware that it had feelings that it experienced emotions and all of that. He was so worried about this and he wanted, you know, I think he made it public by releasing some transcripts of his conversations with it. And I don't think he was allowed to do that under his Google contract, and that was the issue. tThat made a lot of news and Google pushed back and said, no, no, of course it's not sentient. and then there was a lot of debate in the philosophy sphere of what sentient actually means, how you would know if something is sentient. And it Yeah. and it's kind of gone from there.Eric Topol (05:43):Yeah. And then what was interesting is then in March based upon GPT-4 the Microsoft Research Group published this sparks paper where they said, it seems like it has some artificial general intelligence, AGI qualities, kind of making the same claim to some extent. Right?Melanie Mitchell (06:05):Well, that's a good question. I mean, you know, intelligence is one thing, sentience is another. There's a question of whether, you know, how they're related, right? Or if they're related at all, you know, and what they all actually mean. And these terms, this is one of the problems. Of course, these terms are not well-defined, but most, I think most people in AI would say that intelligence and sentience are different. You know something can be intelligent or act intelligently without having any sort of awareness or sense of self or, you know, feelings or whatever sentience might mean. So I think that the sparks of AGI paper from Microsoft was more about this, that saying that they thought GPT-4 four, the system they were experimenting with, showed some kind of generality in its ability to deal with different kinds of tasks. You know, and this, this contrasts with the old, older fashioned ai, which typically was narrow only, could do one task, you know, could play chess, could play Go, could do speech recognition, or could, you know, generate translations. But it, they couldn't do all of those things. And now we have these language models, which seemed to have some degree of generality.The Persistent Gap Between Humans and LLMsEric Topol (07:33):Now that gets us perfectly to an important Nature feature last week which was called the “Easy Intelligence Test that AI chatbots fail.” And it made reference to an important study you did. First, I guess the term ARC --Abstract and Reasoning Corpus, I guess that was introduced a few years back by Francois Chollet. And then you did a ConceptARC test. So maybe you can tell us about this, because that seemed to have a pretty substantial gap between humans and GPT-4.Melanie Mitchell (08:16):Right? So, so, so Francois Chollet is a researcher at Google who put together this set of sort of intelligence test like puzzles visual reasoning puzzles that tested for abstraction abilities or analogy abilities. And he put it out there as a challenge. A whole bunch of people participated in a competition to get AI programs to solve the problems, and none of them were very successful. And so what, what our group did was we thought that, that the original challenge was fantastic, but the prob one of the problems was it was too hard, it was even hard for people. And also it didn't really systematically explore concepts, whether a, a system understood a particular concept. So, as an example, think about, you know, the concept of two things being the same, or two things being different. Okay?(09:25):So I can show you two things and say, are these the same or are they different? Well, it turns out that's actually a very subtle question. 'cause when we, you know, when we say the same we, we can mean sort of the, the same the same size, the same shape, the same color, this, you know, and there's all kinds of attributes in which things can be the same. And so what our system did was it took concepts like same versus different. And it tried to create lots of different challenges, puzzles that had that required understanding of that concept. So these are very basic spatial and semantic concepts that were similar to the ones that Solei had proposed, but much more systematic. 'cause you know, this is one of the big issues in evaluating AI systems is that people evaluate them on particular problems.(10:24):For example, you know, I think a lot of people know that ChatGPT was able to answer many questions from the bar exam. But if you take like a single question from the bar exam and think about what concept it's testing, it may be that ChatGPT could answer that particular question, but it can't answer variations that has the same concept. So we tried to take inside of this arc domain abstraction and reasoning corpus domain, look at particular concepts and say, systematically can the system understand different variations of the same concept? And then we tested this, these problems on humans. We tested them on the programs that were designed to solve the ARC challenges, and we tested them on G P T four, and we found that humans way outperformed all the machines. But there's a caveat, though, is that these are visual puzzles, and we're giving them to GPT-4, which is a language model, a text, right? Right. System. Now, GPT four has been trained on images, but we're not using the system that can deal with images. 'cause that hasn't been released yet. So we're giving the system our problems in a text-based format rather than like, like giving it to humans who actually can see the pictures. So this, this can make a difference. I would say our, our our, our results are, are preliminary .Eric Topol (11:57):Well, what do you think will happen when you can use in inputs with images? Do you think that it will equilibrate there'll be parity, or there still will be a gap in that particular measure of intelligence?Melanie Mitchell (12:11):I would predict there, there will still be a big gap. Mm-hmm. , but, you know, I guess we'll seeThe Biggest Question: Stochastic Parrot or LLM Real Advance in Machine Intelligence?Eric Topol (12:17):Well, that, that's what we want to get into more. We want to drill down on the biggest question of large language models. and that is, are they really you know, what is their level of intelligence? Is it something that is beyond the so-called stochastic parrot or the statistical ability to adjudicate language and words? So there was a paper this week in Nature Human Behavior, not a journal that normally publishes these kind of papers. And as you know it was by Taylor Webb and colleagues at U C L A. And it was basically saying for analogic reasoning ,making analogs, which would be more of a language task, I guess, but also some image capabilities that it could do as well or better than humans. And these were college students. So , just to qualify, they're, they're not, maybe not, they're not fully representative of the species, but they're at least some learned folks. So what did, what did you think of that study?Melanie Mitchell (13:20):Yeah, I found it really fascinating. and, and kind of provocative. And, you know, it, it kind of goes along with a, a many, there's been many studies that have, have been applying tests that were kind of designed for humans, psychological tests to large language models. And this one was applying sort of analogy tests that, that psychologists have done on humans to, to, to large language models. But there's always kind of an issue of interpreting the results because we know these large language models most likely do not think like we do. Hmm. And so one question is like, how are they performing these analogies? How are they making these analogies? So this brings up some issues with evaluation. When we try to evaluate large language models using tests that were designed for humans. One question is, were these tests at all actually in the training data of a large language model? Like, had they, you know, these language models are trained on enormous amounts of text that humans have produced. And some of the tests that that paper was using were things that had been published in the psychology literature.(14:41):So one question is, you know, to what extent were those in this training data? It's hard to tell because we don't know what the training data exactly is. So that's one question. Another question is are these systems actually using analog reasoning the way that we humans use it? Or are they using some other way of solving the problems? Hmm. And that's also hard to tell. 'cause these systems are black boxes, but it might actually matter because it might affect how well they're able to generalize. You know, if I can make an analogy usually you would assume that I could actually use that analogy to understand some new, you know, some new situation by an analogy to some old situation. But it's not totally clear that these systems are able to do that in any general way. And so, you know, I tdo hink these results, like these analogy results, are really provocative and interesting.(15:48):But they will require a lot of further study to really make sense of what they mean, like to when you give, when, when the, the, you know, ChatGPT passes a bar exam, you might ask, well, and let's say it's, you know, it does better than most humans, can you say, well, can it now be a lawyer? Can it go out and replace human lawyers? I mean, a human who passed the bar exam can do that. But I don't know if you can make the same assumption for a language model, because it's the way that it's doing, answering the questions in a way that its reasoning might be quite different and not imply the same kinds of more general abilities.Eric Topol (16:32):Yeah. That's really vital. And something else that you just brought up in multiple dimensions is the problem of transparency. So we don't even know the, the specs, the actual training, you know, so many of the components that led to the model. and so you, by not knowing this we're kind of stuck to try to interpret it. And I, I guess if you could comment about transparency seems to be a really big issue, and then how are we going to ever understand when there's certain aspects or components of intelligence where, you know, there does appear to be something that's surprising, something that you wouldn't have anticipated, and how could that be? Or on the other hand, you know, why is it failing? so what is, is transparency the key to this? Or is there something more to be unraveled?Melanie Mitchell (17:29):I think transparency is, is a big part of it. Transparency, meaning, you know, knowing what data, the system was trained on, what the architecture of the system is. you know, what other aspects that go into designing the system. Those are important for us to understand, like how, how these systems are actually work and to assess them. There are some methods that people are using to try and kind of tease out the extent to which these systems have actually developed sort of the kind of intelligence that people have. So, so one, there was a paper that came out also last week, I think from a group at MIT where they looked at several tasks that were given that GPT-4 did very well on that seemed like certain computer programming, code generation, mathematics some other tasks.(18:42):And they said, well, if a human was able to generate these kinds of thin…

Full show notes at the publisher

Al Gore: The Intersection of A.I. and Climate Change Jun 21, 2023

Show notes

Transcript with some hyperlinksEric Topol (00:00):Hello, Eric Topol here. And what a privilege to have as my guest Al Gore, as we discuss things that are considered existential threats. And that includes not just climate change but also recently the concern about A.I. No one has done more on the planet to bring to the fore the concerns about climate change. And many people think that the 2006 film, An Inconvenient Truth, was the beginning, but it goes way back into the 1980s. So, Al it's really great to have you put in perspective. Here we are with the what's going on in Canada with more than 12 million acres of forest fires that are obviously affecting us greatly, no less the surface temperature of the oceans. And so many other signs of this climate change that you had warned us about decades ago are now accelerating. So maybe we could start off out, where are we with climate change and the climate reality?The Good News on Climate ChangeAl Gore (01:00):Oh, well, first of all, thank you so much for inviting me to be on your podcast again, Eric. It's always a pleasure and especially because you're the host and we, we have very interesting conversations that aren't on the podcast. So, , I'm looking forward to this one. So, to start with climate you know, the old cliche, there's good news and bad news. Unfortunately, there's an abundance of bad news but there's also an awful lot of good news. Let me start with that first and then turn to the more worrying trends. We have seen the passage in the US last August of the largest and most effective best funded climate legislation passed by any nation in all of history. The so-called Inflation Reduction Act is an extraordinary piece of legislation.(01:55): It's billed as allocating $369 billion to climate solutions. But actually, the heavy lifting in that legislation is done by tax credits, most of which are open-ended and uncapped, and a few without any time limits, most a 10-year duration. And the enthusiastic response to the legislation after President Biden signed it has now made it clear that that early estimate of 369 billion is a low-ball estimate, because Goldman Sachs, for example, is predicting that it will end up allocating 1.2 trillion to climate solutions. A lot of other investors and others using economic models are estimating more than a trillion. So, it's really a fantastic piece of legislation and other nations are beginning to react and respond and copy it. One month after that law was passed the voters of Australia threw out their climate denying government and replaced it with a climate-friendly government, which immediately then set about passing legislation that adopts the same goals as the US IRA and the Australian context.(03:19):And they stopped the biggest new coal mine there. And anyway, one month after that, in October, the voters of Brazil threw out their former president often called the “Trump of the Tropics” and replaced him with a new president, a former president who's a new president, who has pledged to protect the Amazon and the European Union in responding to the evil, evil and cruel invasion of Ukraine by Russia. And the attempted blackmail of nations in Europe, dependent on Russian gas and oil responded not by bending their knee to Vladimir Putin, but by saying, wait a minute, this makes renewable energy, freedom, energy. And so they accelerated their transition. And so these are all excellent signs and qualifies as good news. The other good news is not all that new, but it's still continuing to improve.(04:28):And that is the astonishing reductions in cost for electricity produced by solar and wind, and the reductions in cost for energy storage, principally in batteries and electric vehicles and a hundred other less well known technologies that are extremely important. We're in the midst of early stages of a sustainability revolution that has the magnitude of the industrial revolution, coupled with the speed of the digital revolution. And we're seeing it all over the place. It’s really quite heartening. One quick example last, the, the biggest single source of global warming pollution is the generation of electricity with gas and coal. Well, last year, if you look at all the new electricity generation capacity installed worldwide 90% of it was renewable. In India, 93% was solar and wind. And India's pledged not to give permits for any new coal burning plants for at least five years, which means never, probably because this cost reduction curve, as I mentioned, is still continuing downward electric vehicles, we're now seeing that the purchases have reached 15% of the market globally.(05:56):Norway's already at 50%. They've actually outlawed the sale of any new internal combustion engines. And indeed, many national and even municipal and state jurisdictions have prospectively served notice that they, you won't be able to buy them after a certain day, 2030, in many cases and the auto companies and truck and bus companies have long since diverted their research money all their R & D is going into EVs now. And that's the second largest source of global warming pollution. I could go through the others, but I want, I'll just tell you that there is a lot of good news.And the Bad NewsNow, the bad news is we're still seeing the crisis get worse, faster than we're deploying all of these solutions. And, the inertia in our political and economic systems is partly a direct result of huge amounts of lobbying and campaign contributions and the century old net of political and economic influence built up by the fossil fuel industry.(07:18):And they're opposing every single solution at the state level, the local level, the national level, the international level. Now, this COP 28 [the 2023 United Nations Climate Change Conference] coming up at the end of the year in the United Arab Emirates is actually chaired by an oil and gas company CEO-- It's preposterous. And they already have in the last two COPS, more lobbyists registered as participants than all than the five or six largest national delegations combined. And we're seeing them really oppose this change. And meanwhile, the manifestations of the crisis are steadily worsening. You mentioned the fires in Canada that are predicted to burn all summer long. And I was in New York City last week, and you, you know, from the news stories it, it was horrific. I got there the day after the worst day, oh my God.(08:21):But I saw and heard from people just the tremendous problems that people have. It's also going on in Siberia, by the way, and these places that are typically beyond the reach of TV crews and networks that don't capture our attention unless something happens to blow the smoke to where we live. And that's what's happened here. But there are many other extremely worrying manifestations that aren't getting much attention. I do think we're going to solve this, Eric. I'm very optimistic, but the question is whether we will solve it in time. We are what's the right way to say this? We're tiptoeing through a minefield with tripwires and toward the edge of a cliff. I don't want to torture the metaphor, but actually there are several extremely dangerous threats to ecological systems that are in a state of balance now, and are being pushed out of their equilibrium state into a different format.(09:35):The ocean currents--we're already seeing it with the jet stream in the northern hemisphere. You may have seen on the weather maps. They're now using these a lot where it's getting loopier and more disorganized. That's what the last few winners has, has pulled these big loops, have pulled arctic air down into areas far south in the US and in other regions, by the way. And it’s making a lot of the extreme events worse. Now, we're entering an El Nino phase in the Pacific Ocean comes around every so often, and this one is predicted to be a strong one, and that's going to accentuate the temperature increase. You know, it was [recently] 110 degrees last week in Puerto Rico, 111 degrees in several countries in Southeast Asia.(10:31):Last summer, China had a heat wave that the historians say about, which the historians say there's nothing even minimally comparable in all prior known, and the length, the extent, the duration, the intensity. And we saw monsoons lead to much of Pakistan underwater for an extended period of time. I could go on, but the net and balance out the good news and the bad news we are gaining momentum. And soon we are going to be gaining on the crisis itself and start deploying solutions faster than it's getting worse. So I remain optimistic, and I always remind people, if you doubt we have the political will to see this through, remember that political will is itself a renewable resource.The Intersection of A.I. and Climate ChangeEric Topol (11:27):Yeah, that's a great optimistic point, and we sure appreciate that, because it's pretty scary to see these trends that you reviewed. Now, as you know recently there was a large group of AI scientists this one led by Sam Altman of OpenAI, who put out a statement, a one-sentence statement, and it said, “Mitigating the risk of distinction from ai, which you and are enthusiastic about, should be a global priority alongside other societal scale risks, such as pandemics and nuclear war.” Well, obviously, also climate change. So how do you see the AI intersection of climate change? Because as you well know, GPT-4, having pre-trained with some 30,000 graphic processing units [GPUs], the issues about consumption of energy carbon emissions, the need for water cooling, is AI going to make this situation worse, or will it make it better?Al Gore (12:33):Well, yeah. You know, I understand. Well, both would be my answer. And we don't have enough data yet to really know for sure which way it will tip. Maybe we'll talk about the existential risks from generative AI. As this conversation continues, there are many who have spoken up and said, well, wait a minute, before we focus on that, we need to look at the risks that are right, staring us right in the face. I mean, the use of these AI driven algorithms, not necessarily generative AI, but the AI-driven algorithms in social media are causing tremendous harm right now. You've heard about the rabbit holes that people get drawn down into on the internet. That's because of the AI-driven algorithms and the tracking of confidential information about what people are looking at and what they're interested in.(13:40):And these are rabbit holes are ,a little bit not to shift metaphors, a little bit like pitcher plants in that they have slippery slides and, oh, and, you know, what's at the bottom of the rabbit hole? That's where the echo chamber is. And when you spend long enough in the echo chamber, then those who are feeding the information to you weaponize a new form of AI, not artificial intelligence, artificial insanity. And, and we see it all over the place where people are utterly convinced of completely ridiculous and provably false conclusions and, and motivated to go out and act in the real world. On that basis we, we see the fakes and the concerns about video and audio deep fakes, and how that's going to have an impact on us and, and all manner of other concerns that need to need to be addressed.(14:43):But the existential threat is one that I do want to come back to. But, turning to your specific focus on whether it is going help or hurt or both where climate is concerned, I have co-founded a coalition called Climate TRACE that uses AI in an extremely effective, beneficial way. Trace stands for tracking real-time atmosphere, carbon emissions, and we have a coalition of AI firms, NGOs, university groups and the whole coalition works together to identify with AI, the point source of every single significant stream of emissions of global warm inclusion everywhere on the planet. We released it at the last United Nations Conference, the one that was held in Egypt last year. The top 72,000 emission point sources around the world this fall; we will release the top 70 million emission sources.(15:54): We also have every agricultural field in the world down to a 10 meter by 10 meter resolution. We have all, every single power plant, all the steel mills, every large ship, every large plane, most every well, we have all of the significant greenhouse gas emissions that wouldn't have not, that would not have been possible without ai. Now, this is not generative AI. We have used generative ai --not ChatGPT--we tried that, but there are others that are actually more proficient in the views of our team members at writing code. It has saved us time and enhanced our productivity in writing code. So that's one example where AI has been a big help. And we see it in modeling, and we see it in the preparation for adaptation and in other ways. Now, the downside is, you said in your introductory phrasing that the energy requirements and the emissions are just enormous because it is an extremely energy intensive exercise.(17:09): And you have to have the GPUs as well as the energy. So it's you could call it “oligopogenic”-- that may not be a word. It may be a hallucination, like GPT is famous for, but what I mean is it, it does tend to favor a very small number, a very wealthy, very powerful, very large companies. Basically, Google and Microsoft are driving the, the rest of the world to try to desperately catch up. You know, the CEO of Microsoft. They stole a march on Google with the release of ChatGPT and then that fascinated people and the pickup and use of GPT unbelievable is just, it, it's there's been nothing like it in.(18:19):Previous technological history. The CEO said that he wanted to make call Google out and make him dance. Well you know, Peggy Noonan said in one of her columns, that's not a responsible way for the CEO of such a company to talk. I, I like him, and I'm not really taking a poke at him, insofar as I'm making the point that there're really two companies, and the internal dynamic between the two is driving this frenzy of investment and activity, and the underlying platform, the large language models, they're all almost a commodity now. They're all over the place and have been for a while. But the need for the GPUs, the need for the energy consumption that's limiting the cutting edge developments to these two companies. For now, China doesn't trust it because they don't trust the enhanced political influence.(19:22):It might give those using it or the enhanced insight. And there are others that will try to find a way to use it, of course. But the, the emissions itself are extremely harmful and the use of generative AI in the hands of irresponsible actors. And, unfortunately, we're human beings and we have a lot of irresponsible actors around this, around this country, around the world. And they could use that to really put climate disinformation into high gear. They, they can use it in a variety of ways to further enhance the disruption, the disruptive tactics they've used in the past.Eric Topol (20:15):Yeah. Well, that's what I wanted to get into more on this. We have, I think, you know, if you want to put an existential risk at the highest level, maybe if you were assign 10 to climate change and you've brought up the fact that the large language models generative AI will make worse, the things we've already seen, the, the hacking of democracy and all the fake stuff that's the conspiracy theories that it will reinforce. And the question is, where are you, where did you place the whole generative AI era that we've now…

Full show notes at the publisher

Hannah Davis: A 360° on Long Covid Jun 06, 2023

Show notes

TRANSCRIPTEric Topol (00:00):Hello, this is Eric Topol, and it's really a delight for me to welcome Hannah Davis who was the primary author of our recent review on Long Covid and is a co-founder of the Patient-Led Research Collaborative. And we're going to get into some really important topics about citizen science, Long Covid and related matters. So, Hannah, welcome.Hannah Davis (00:27):Thank you so much for having me.Eric Topol (00:29):Well, Hannah, before we get into it I thought because you had a very interesting background before you got into the patient led research collaborative organization with graphics and AI and data science. Maybe you could tell us a bit about that.Hannah Davis (00:45):Sure. Yeah. Before I got sick, I was working in machine learning with a particular focus on generative models for art and music. so I did some projects like translating data sets of landscapes into emotional landscapes. I did a project called The Laughing Room, where there was a room and you went in and the room would listen to you and laugh if it thought you said something funny, . and then I did a lot of generative music based on sentiment. So I, I did a big project where I was generating music from the sentiment of novels and a lot of kind of like critical projects, looking at biases in data sets, and also curating data sets to create desired outcomes in these generative models.Eric Topol (01:30):So, I mean, in a way again, you were ahead of your time because that was before ChatGPT in November last year, and you were ahead of the generative AI curve. And here again, you're way ahead in in the citizen science era as it particularly relates to the pandemic. So, I, I wonder if you could just tell us a bit I think it was back, we go back to March, 2020. Is that when you were hit with Covid?Hannah Davis (01:59):Yes.Eric Topol (02:00):And when did you realize that it wasn't just an acute phase illness?Hannah Davis (02:06): for me, honestly, I was not worried at all. I, my first symptom was that I couldn't parse a text message. I just couldn't read it, thought I was tired. an hour later, took my temperature, realized I had a fever, so that's when I kind of knew I was sick. but I really just truly believed the narrative I was going to get better. I was 32 at the time. I had no pre-existing conditions. I just was, you know, laying around doing music stuff, not concerned at all. And I put a calendar note to donate plasma two weeks out, and I was like, you know, I'm going to hit that mark. I'm going to donate plasma, contribute, it'll be fine. And that day came and went. I was still, you know, pretty sick with a mild case. You know, I didn't have to be hospitalized.(02:49):I didn't have severe respiratory symptoms. but my neurological symptoms were substantial and did increase kind of over time. And so I, I was getting concerned. Three weeks went by, still wasn't better. And then I read Fiona Lowenstein’s op-ed in the New York Times. They were also very young. They were 26 at the time, they had been hospitalized, and they had this prolonged recovery, which we now know as Long Covid. and they started the Body Politic Support Group joined that saw thousands of people with the same kind of debilitating brain fog, the same complete executive functioning loss, inability to drive, forgetting your family members' names who were all extremely young, who all had mild cases. and that's kind of when I got concerned because I realized, you know, this was not just happening to me. This was happening to so many people, and no one understood what was happening.Eric Topol (03:49):Right. extraordinary. And, and was a precursor, foreshadowing of what was to come. Now, here it is, well over three years later. And you're still affected by all this, right?Hannah Davis (04:02):Yes. Pretty severely.Eric Topol (04:04):Yeah. And I learned about that when I had the chance to work with you on the review. You were the main driver of this review, and I remember asking you, because I, I didn't know anyone in the world that was tracking Long Covid like you and to be the primary author. And then you sent this outline, and I had never seen an outline in all my years in academic medicine. I never saw an outline like this of the review. I said, oh my God, this is incredible. So I know that during that time when we worked on the review together, along with Lisa McCorkell and Julia Moore Vogel, that, you know, there, there were times when you couldn't work on it right there, there were just absolutely, you would have some good days or bad days. And, and that's the kind of, is that kind of the way is, how it goes in any given unit time?Hannah Davis (04:55):I think generally, I, I communicated as like 40% of my function is gone. So, like, I used to be able to have very, very full days, 12 hour days would work, would socialize, would do music, whatever. you know, I, I have solidly four functional hours a day. on a good day, maybe that will be six. On a bad day, that's zero. And when I push myself by accident, I can get into a crash that can be three to seven days easily. Hmm. and then I'm, then I'm just not, you know, able to be present. I don't feel here. I don't feel cognitively able, I can't drive. And then I'm just completely out of the world for a bit of time.Eric Topol (05:35):Yeah. Wow. So back in the early days of when you were first got sick and realized that this was not going to just go away, you worked with others to form this Patient -Led Research Collaborative organization, and here you are, you didn't have a medical background. You certainly had a data science and computing backgrounds. But what were your thoughts? I mean, citizen science has taken on more of a life in recent years, certainly in the last decade. And here there's a group of you that are kind of been leading the charge. we'll get to, you know, working with RECOVER and NIH in just a moment. But what were your thoughts as to whether this could have an impact at working with these, the other co-founders?Hannah Davis (06:27):I think at first we really didn't realize how much of an impact we were going to have. The reason we started collecting data in the first place really was to get answers for ourselves as patients. You know, we saw all these kind of anecdotes happening in the support group. We wanted to get a sense of which were happening the most at what frequency, et cetera. and it really wasn't until after that when like the CDC and WHO started reaching out, asking for that data, which was gray literature at the time that we kind of realized we needed to formalize this and, and put out an official paper which was what ended up being the second paper. But the group that we formed really is magical, I think like, because the primary motivator to join the group was being sick and wanting to understand what was happening. And because everyone in the group only has the kind of shared experience of, of living with Long Covid, we ended up with a very, very diverse group. Many, many different and I think that really contributed to our success in both creating this data, but also communicating and, and doing actionable policy and advocacy work with it.Eric Topol (07:42):Did you know the folks before? Or did you all come together because of digital synapses?Hannah Davis (07:47):Digital synapses? I love that. Absolutely. No, we didn't know each other at all. they're now all, you know, they're my best friends by far. you know, we've been through this, this huge thing together. but no, we didn't meet in person until just last September, actually. And many of them we still haven't even met in person. which makes it even more magical to me.Eric Topol (08:13):Well, that's actually pretty extraordinary. So together you've built a formidable force to stand up for the millions and millions of people. As you wrote in the review, 65 million people around the world who are suffering in one way or another from Long Covid. So just to comment about the review --you know, I've been working in writing papers for too long, 35 years. I've never, in my entire career, over 1300 peer reviewed papers on varied topics, ever had one that's already had 900,000 downloads, is the fourth most cited paper and Altmetric since published the same timeframe in January of all 500,000 peer-reviewed papers. Did you ever think that the, the work that, that you did and our, you know, along with Lisa and doing, and I would ever have this type of level of interest?Hannah Davis (09:16):No, and honestly, it's so encouraging. Our, our second paper to me did very well. and, you know, was, was widely viewed and widely cited, and this one just surpassed that by miles. And I think that it's encouraging because it communicates that, that people are interested, right? People, even if they don't understand what long covid is, there is a huge desire to know. And I think that putting this out in this form, focusing on the biomedical side of things really gives people a, a tool to start to understand it. And from the patient side of things, more than any other paper I've heard we, we get so many comments that are like, oh, I brought this to my doctor and, you know, the course of my care change. Like he believed me and he started X treatment. and that, that's the kind of stuff that just makes us so, so meaningful. and I'm so, so grateful that, that we were able to do this.Eric Topol (10:16):Yeah. And as you aptly put it, you know, a work of love, and it was not easy because the reviewers were not not all of them were supportive about the real impact, the profound impact of long covid. So when you now every day you're keeping track of what's going on in this field, and there's something every single day. one of the things, of course is that we haven't really seen a validated treatment all this time, and you've put together a list of candidates, of course, it was in the review, and it constantly gets revised. What are some of the things that you think are alluring from preliminary data or mechanisms that might be the greatest unmet need right now of, of getting some relief, some remedy for this? What, what, what's your sense about that?Hannah Davis (11:13):I think the one I'm most excited about right now are JAK/STAT inhibitors. And this is because one of the leading researchers in viral onset illness Ron Davis and that group believe that basically they're, they have a shunt hypothesis, and that means they, they basically think there's a switch that happens in the body after you've, you've had a viral illness like this, and that that switch can actually be unswitched. And that, to me, as a patient, that's very exciting because, you know, that that's what I imagine a cure kind of looks like. and they did some computational modeling and, and identified JAK/STAT inhibitors as one of the promising candidates. so that's from like the, like hypothetical side that needs to be tested. And then from the patient community, from some things we're seeing I think really easily accessed ones include chromolyn sodium.(12:14):So these are prescription antihistamines. they're both systemic. So Coen has been seeming to work for patients with brain fog and sleep disorders. And chromolyn sodium particularly works in, in patients with gastrointestinal mast cell issues. People are going on to kind of address the micro clots. I, for me personally, has been one of the biggest changers game changers for my brain fog and kind of cognitive impairment type things. but there's so many others. I mean, I think we, we really wanna see trials of anticoagulants. I'm personally really excited to start on ivabradine which is next up in my queue. And, and seems to have been a, a game changer for a lot of patients too. I V I G has worked for patients who are, have been able to get it, I think for both I V I G and ivabradine. Those are medications that are challenging to get covered by insurance. And so we're seeing a lot of those difficulties in, in access with a couple of these meds. But yeah, just part of, part of the battle, I guess,Eric Topol (13:32):You know, one of the leading of many mechanisms that in this mosaic of long covid is the persistence of virus or virus components. And there have been at least some attempts to get some Paxlovid trials going. Do you see any hope for just dealing, trying to inactivate the virus as a way forward?Hannah Davis (13:54):Absolutely. Definitely believe in the viral persistence theory. I think not only Paxlovid, but other a covid antivirals. I know that Steve deas and Michael Paluso at U C S F are starting a couple long covid trials with other covid antivirals that yeah, for sure. I think they all obviously need to be trialed A S A P. And then I also think on the viral persistence lens, ev like almost everyone I know has viral reactivation of some sort like EBV, CMV, VZV, you know, we obviously see a lot of chickenpox or shingles reactivations and antivirals targeting those as well I think are really important.Eric Topol (14:41):Yeah. Well, and I also, just the way you're coming out with a lot of this, you know terminology and, you know science stuff like I V I G for intravenous immunoglobulin and for those who are not, you know, just remember, this is a non-life science expert who now has become one. And that goes back again to the review, which was this hybrid of people who had long covid with me who didn't to try to come up with the right kind of balance as to, you know, what synthesizing what, what we know. And I think this is something the medical profession has never truly understood, is getting people who are actually affected and, and becoming, you know, the real experts. I mean, I, I look to you as one of the world's leading authorities, and I learn from you all the time.(15:35):So that goes to RECOVER. So there was a long delay in the US to recognize the importance of long covid. Even the UK was talking to patients well before they ever had a meeting here in the us, but eventually, somehow or other they allocated a billion dollars towards long covid research at the NIH. And originally, you know, fortunately Francis Collins, when he was director, saw the importance, and he, I learned bequeathed that 2 of the NIH institutes, one of the directors, Gary Gibbons visited me recently because of a negative comment I made about RECOVER. But before I go over my comment, you've been as he said, you, and Lisa McCorkell ,among others from the Patient-led Collaborative have had a seat at the table. That's a quote from Gary. Can you tell us your impression about RECOVER you know, in terms of at least they are including Patient-Led research folks with long covid as to are they taking your input seriously? And what about the billion dollars ?Hannah Davis (16:46):Oh, boy. tricky question. I don't even know where to start. Well, I mean, so I think recover really messed up by not putting experts in the field in charge, right? Like we are, we have from the beginning have needed to do medical provider education at the same time that all these studies started getting underway. And that was just a massive amount of work to try to include the right test to convince medical professionals why they weren't necessary. all that could have been avoided by putting the right people in charge. And unfortunately, that didn't happen. unfortunately recovers our, our best hope still or at least the, the best funded hope. so I really want to see it succeed. I think that they, they have a long way to go in terms of, of really understanding why patient representation matters and, and pat…

Full show notes at the publisher

Peter Lee and the Impact of GPT-4 + Large Language AI Models in Medicine May 22, 2023

Show notes

Link to the book: The AI Revolution in MedicineLink to my review of the bookLink to the Sparks of Artificial General Intelligence preprint we discussedLink to Peter’s paper on GPT-4 in NEJMTranscript (with a few highlights in bold of many parts that could be bolded!)Eric Topol (00:00):Hello, I'm Eric Topol, and I'm really delighted to have with me Peter Lee, who's the director of Microsoft Research and who is the author, along with a couple of colleagues for an incredible book called The AI Revolution in Medicine, GPT-4 and Beyond. Welcome, Peter.Peter Lee (00:20):Hello Eric. And thanks so much for having me on. This is a real honor to be here.Eric Topol (00:24):Well, I think you are in the enviable position of having spent now more than seven months looking at GPT-4’s S capability, particularly in the health and medicine space. And it was great that you recorded that in a book for everyone else to learn because you had such a nice head start. I guess what I wanted to start with is, I mean, it's, it's a phenomenal book. I [holding the book up], this prop. I can't resistPeter Lee (00:52):Eric Topol (00:53):When, when I got it, I, I couldn't, I stayed up most of the night because I couldn't put it down. It was, it is so engrossing. But when you, when you first got your hands on this and started testing it, what were, what were your initial thoughts?Peter Lee (01:09):Yeah. I, let me first start by saying thank you for the nice words about the book, but really, so much of the credit goes to the co-authors, Carey Goldberg and Zach Kohane and Corey in particular took my overly academic writing. I suspect you have the same kind of writing style as well as Zach's pretty academic writing and helped turn it into something that would be approachable to non-computer scientists and as she put it, as much as possible as a page turner. So I'm glad that her work helped make the, the book an easy read. I,Eric Topol (01:54):I want to just say you're very humble because the first three chapters that you wrote yourself were clearly the, the best ones for me. Anyway. I don't mean to interrupt, but it, it, it is an exceptional book, really.Peter Lee (02:06):Oh thank you very much. It means a lot. Hearing that from you. You know, my own view is that the, the best writing and the best analyses and the best ideas for applications or not of this type of technology in medicine are yet to come. But you're right that I did benefit from this seven-month head start. And so, you know, I think the timing is, is very good. but I'm hoping that much better books and much better writings and ideas will come, you know, when you start with something like this, I, I suspect, Eric, you had the same thing. you start off with a lot of skepticism and I, in fact, I sort of now made light with this. I talk about the nine stages of grief that you have to go through.(02:55): I was extremely skeptical. Of course, I was very aware of GPT 2, GPT 3 and GPT 3.5. I understand, you know, what goes into those models really deeply. and so some of the claims, when I was exposed to the early development, GPT-4 just seemed outlandish and impossible. So I, I was, you know, skeptical, somewhat quietly skeptical. We've all been around the block before and, you know, we've heard lots of AI claims and I was in that state for maybe more than two weeks. And then I started to become in that two weeks annoyed, because I know that some of my colleagues like falling into what I felt was the trap of getting fooled by this technology. And then that turned into frustration and fear. I actually got angry. And one colleague who I won't name I've since had to apologize because then I into the phase of amazement because you start to encounter things that you can't explain that this thing seems to be doing that turns into joy.(04:04): I remember the exhilaration of thinking, wow, I did not think I would live long enough to see a technology like this. and then intensity, There was a period of about three days when I didn't sleep, I was just experimenting. Then you run into some limits and some areas of puzzlement and that's a phase of chagrin. And then real dangerous missteps and mistakes that this system can make that you realize might end up really hurting people. and then, you know, ChatGPT gets released and to our surprise it catches fire with people. And we learn directly through communications that some clinicians are using it in clinical settings. And that heightens the concern. And I, I can't say I'm in the ninth stage of enlightenment yet, but you do become very committed to wanting to help the medical community get up to speed and to be in a position to take ownership of the question of whether, when, and how a technology like this should be used. and that was really the motivating force behind the book. And it, it was really that journey. And that journey also has given me patience with everyone else in the world, because I realize everyone else in the world has to go through those same nine, nine stages.Eric Topol (05:35):Well, those stages that you went through are actually a great way to articulate this pluripotent technology. I mean, I think you, you touched on that chat. ChatGPT was released November 30th and within 90 days had a billion distinct users, which is beyond anything in history. And then of course, this transcended that quite a bit as you showed in the book coming out in you know, just a very short time in March. right. And I think a lot of people want access to GPT-4 because they know that there is this jump in its capabilities. But the book starts off after Sam Altman's forward, which was also nice because he said, you know, this is just an early, as you pointed out there, there's a lot more to come in the large language model space.(06:30):But the grabber to me was this futuristic, this second year medical resident who's using an app on the phone to get to the latest GPT to help manage her patient, and then all the other things that it's doing to check on her patients and do all the things that are the tasks that clinicians don't really want to do, that they need help with. And that just grabs you as to the futuristic potential, which may not be so far away. And I think then you get into the nuts and bolts, but one of the things that I think is a misnomer that you really nailed is how you say it isn't just that it generates, but it really is great at editing and analyzing. And here it's, it's called generative AI. Can you, can you expound on that? And it's unbelievable conversationalist capability.Peter Lee (07:23):Yeah. you know, the term Generative AI, I tried for a while to push back on this, but I think it's just caught on and I've given up on that. And I get it. You know, I, I think especially with ChatGPT it's of course reasonable for the public to be, you know infatuated with a thing that can write love letters, write poetry and that generative capability. and of course, you know school children writing their essays and so on this way. But as you say one thing we have discovered through a lot of experimentation is it's actually somewhat of a marginal generator of text. I would not say at all. That is, it is not as good a poet as good human poets. It's not the, you know, people have programmed GPT-4 to try to write whole novels and it can do that,(08:24):they aren't great. and it's a challenge, you know within Microsoft, our Nuance division has been integrating GPT-4 to help write clinical and encounter notes. and you can tell it's just hitting at the very limits of the capabilities in and of the intelligence of GPT-4 to be able to do that well. But one area where it really excels is in evaluating or judging or reviewing things. And we've seen that over and over again. in chapter three. You know, I have this example of its analysis of some contemporary poetry which is just stunning in its kind of insights and its use of metaphor and allegory. And but then in other situations in interactions with the New England Journal Journal of Medicine experimentations with the use of GPT-4 as an adjunct to the review process for papers it is just incredibly insightful in spotting inconsistencies missing citations to precursor studies to understanding lack of inclusivity and diversity, you know, in approach or in terminology.(09:49):And these sorts of review things end up being especially intriguing for me when we think about the whole problem of medical errors and the possibility of using GPT-4 to look over the work of doctors, of nurses of insurance, adjudicators and others, just as a second set of eyes to check for errors check for kind of missing possibilities if there's a differential diagnosis. Is there a possibility that's been something that's been missed? If there's a calculation for an IV medication administration, well, it's a calculation done correctly or not. And it's in those types of applications of GPT-4 as a reviewer, as a second set of eyes that I think I've been especially impressed with. And we try to highlight that in the book.Eric Topol (10:43):Yeah. That's one of the very illuminating things about going well beyond what are the assumed utilities in a little bit, we'll talk about the liabilities, but certainly these are functions part of that flurry potent spectrum that I think a lot of people are not aware of. One, particularly of interest in the medical space is something I had not anticipated as, you know, when I wrote a Deep Medicine chapter, “Deep Empathy,” I said, well, we got to rely totally on humans for that. But here you had examples that were quite stunning of coaching physicians by going through their communication, their note and saying, you know, you could have been more sensitive with this. You could have done this, but you, you could be more empathic. And as you know, since the book was published, there was an interesting study that compared a couple hundred questions directed to physicians and then to ChatGPT, which of course isn't necessarily called, we wouldn't say it's state of the art at this point, right. But what was seen that chatbot exhibited, the more empathy, the more sensitive, higher quality responses. So do you think, ultimately that this will be a way we can actually use technology to foster a better communication between clinicians and patients?Peter Lee (12:10):Well I'll try to answer that, but then I want to turn the question to you because I'm just dying to understand how others especially leading thinkers like you think about this. Because as a human being and as a patient, there's something about this that doesn't quite sit right. You know I, I want the empathy to come from my doctor, my human doctor that's in my heart the way that I feel. And yet there's just no getting around the fact that GPT-4 and even weaker versions like GPT 3.5, CHatGPT can be remarkably empathetic. And as you say, there was that study that came out of UC, San Diego Medicine, Johns Hopkins Medicine that you know, was just another fairly significant piece of evidence to that point.Here's another example. You know, my colleague Greg Moore was assisting a patient who had late stage pancreatic cancer.(13:10):And there was a real struggle for both the specialists and for Greg to know what to say to this desperate patient how to support this patient. And the thing that was remarkable Greg decided to use GPT-4 to get advice and they had a conversation and there was very detailed advice to Greg on what to say and how to support this patient. And at the end when Greg said, thank you, GPT-4 said, and you're welcome, Greg, but what about you? You know, do you have all the support that you need? This must be very difficult for you. So the empathy just goes remarkably deep. And, you know, if you just look at how busy good doctors and especially nurses are, you can start to realize that people don't necessarily have the time to think about that.(14:02):And also that what GPT-4 is suggesting ends up being a prompt to the human doctor or the human nurse to actually take the time to reflect on what the patient might need to hear, right. What might be going through their minds. And, and so there is some empathy aid going on here. At the same time, I think as a society, we have to understand how comfortable we are with the idea of this concept of empathetic care being assisted by a machine. and this is something that I'm very keen and curious about just in the medical community. And, and that's why I wanted to turn the question back around to you. how do you see this?Eric Topol (14:46):Yeah, I didn't foresee this, but I, and I also recognize that we're talking about a machine vector of it. I mean, it's a pseudo-empathy of sort. But the fact that it can process where it can be improved and it can help foster essentially are features that I think are extraordinary. I, I wouldn't have predicted that. And I've seen now, you know, many good examples in the book and, and even beyond. So it's a welcome thing and it adds another capability which is partly isn't that, that physicians and nurses are lacking empathy, but because their biggest issue, I think is lacking time. Yes. And the fact that someday there's a rescue in the works, hopefully, that a lot of that time of tasks that are, you know, the data clerk functions and other burdens right, will be alleviated the keyboard liberation that has been a fantasy of mine for some years, maybe ultimately will be achieved.(15:52):And the other thing I think that's really special in the book that I wanted to comment, there is a chapter by I think Carey Goldberg. And that was about the patient side, right? And this is what we, we all, the talk is about, you know, doctors and clinicians, but it's the patients who could derive the most. And out of those first billion people that used ChatGPT, many were of course health and medical question conversations. But these are patients, we're all patients. And the idea that you could have a personal health advisor, a concept which was developed in that chapter, and the whole idea that that as opposed to a search today, that you could get citations and it would be at the, at the literacy level of the person asking them, making the prompts. Yeah. Could you comment about that? Because that seems to be very much underemphasized, this democratization of this high level capability of getting you know, very useful information and conversation.Peter Lee (16:56):Yeah. And I think also this is also where some of the most difficult societal and regulatory questions might come, because while the medical community knows how to abide by regulations, and there is a regulatory framework, the same is much less true for a doctor in your pocket, which is what GPT-4 and, you know, other large language models that are emerging can, can become. And you know, I think for me personally I have come to depend on GPT-4. I use it through the Bing search engine. sometimes it's simple things that previously weren't mysterious. Like I received an explanation of benefits notice from my insurance company, and it is this notice it has some dollar figures in it. It has some CPT codes, and I have no idea. And sometimes it's things that my son or my wife got treated for.(17:55):It's, it's just mysterious. It's great to have an AI that can decode these things and can answer questions. similarly, when I go for a medical checkup and I get my blood test results just decoding those CBC lab test numbers, it, it's, again, something that is just incredible convenience. But then even more you know, my fat…

Full show notes at the publisher

Straight talk with Michael Osterholm May 05, 2023

Show notes

Transcript Eric (00:00):Okay. Hello, this is Eric Topol and this is a rare privilege for me to interview my favorite epidemiologist, Dr. Michael Osterholm. He is the Regents Professor of the University of Minnesota. He's director of CIDRAP, which is certainly one of the leading entities around the world for public health. And, we've been friends for the last few years, which we'll we'll talk about. So, welcome Michael. Such a great privilege to have you today.Michael (00:31):Well, thank you, the honor, really is mine. As I have shared with you and others know very well--you have been a real mentor to me and many others during this pandemic. And, I could never repay you adequately for all that you've helped teach me throughout these last three years. It's been immeasurable.Eric (00:49):No, if you're too kind, I think it's much different. The opposite way. I've learned so much from you because this isn't my area, as you well know. I thought we'd start with, of course, right now things are relatively good for the pandemic in the United States and mostly around the world, with relatively less cases, less hospitalizations and deaths. But obviously still people are getting infected. And maybe you can tell us about the recent case that you went through that would be enlightening.End of the Pandemic?Michael (01:28):Yeah, I think we're all trying to understand when the pandemic ends. And, as we've discussed many times before, we'll probably know that about a year after it ends, then we'll say, yep, that was the end of it. Don’t for a moment think that at the end means that there won't be cases. You know, for every infectious agent that we think of when causing a pandemic, they still come back, whether it be influenza, or potentially coronaviruses. They will, they will continue to circulate. It's a matter of how many cases occur, how many people die. And I think that's an important point. There isn't really a definition for when a pandemic ends. It's, I guess it's just when you feel like it's over. And clearly the world has come to that conclusion already. You don't need a, an epidemiologist or a politician to tell 'em that the pandemic's over that they feel that we're still seeing about 165 deaths a day in this country from Covid.(02:24):So it's hardly gone away completely. But we do have to acknowledge it. Most of those deaths are older individuals, people who have not been vaccinated recently with bivalent boosters. And in that regard, we could surely even reduce the illnesses further. I don't have any faith right now in the surveillance systems that have been set up to look at cases around the world. We've pretty much dismantled that. We are not testing people that we results in reports being made to public health agencies, whether in this country or anywhere else in the world. So I really look at two other things. One is deaths. And even they're realizing that still is a challenge in terms of how complete death reporting is due to covid. But then the other thing we're looking at, which has been really, you might say, public health revolution during the pandemic, and I say revolution cause it's really changed things.(03:19):And that is the issue of wastewater surveillance. And we've been able to ascertain in many areas of the world, in fact, with using wastewater surveillance, a much better sense of how much virus is in the community. And so, just in following with your very thoughtful comment about case numbers dropping, that's exactly what we're seeing in most locations in this country too. We, for example, here in the Minneapolis St. Paul area, have seen a dramatic decrease in wastewater activity in the last two months. So I think we're in a place right now where I can hope it'll only get better. On the other hand, you know, I have a lot of respect for this virus, and frankly, we all ought to have a lot of humility. We don't know if another variant will emerge that with, given how much immunity we have in our population will somehow break through that and cause increase in surgeon cases or whether this will become kind of the norm and we'll see less and less.On Getting Covid(04:16):Now, you asked me about my case. Yeah. I have to say that, I speak about this with, with really some trepidation in the sense that I was not gonna get this. I had and very faithful throughout the course of the pandemic, where in my N 95 respirator when I went out and about, I had been fit tested. In addition, when we finally did socialize in our home, we had a, what became affectionately known as the Osterholm Home Rule. You could not have had known contact with someone of the, with Covid in the five previous days. You could have no symptoms yourself on the day of, and you had to test negative bilateral flow test within three to four hours of coming. And we would entertain small four, the six party, parties, and it was going wonderful.(05:07):And then on March 10th, the night of March 10th, a colleague from work came over with Fern and myself. Three of us had dinner. We went down our elevator in our building here, which were 31 stories up. No one else is in the elevator. And then we proceeded to go to a very small music venue where we wore N95s. We were some distance from any other people, and we were there for an hour and 45 minutes. And, literally two days later, almost 48 hours later, all three of us developed symptoms. None of us converted for another 24 hours. And then at that point, we all three tested lateral flow positive.. We all three took Paxlovid. I took it and was starting to feel better after that fifth day.(05:59):And then I kind of crashed and at that point, I got a second, , five day course of Paxlovid and started to feel better. And, I'm you know, was very happy to have this behind me. However, over the course of the last 10 days, I have really had significant fatigue. You know I'm not one that sleeps a lot But, I can tell you there are multiple times in a day where I'm doing something like even doing what I'm doing right now where I just feel like I just need to fall asleep. It’s been really a challenge. The other thing that happened, which was in retrospect a little bit more concerning than I realized at the time, there was a period at about day 10 to 14 into my illness, I started losing my memory on many, many things of, you know, importance.(06:53):I couldn't, for example, tell you what was that drink that is: a champagne, orange juice combination. I couldn't find the word mimosa if my life depended on it. If somebody asked me who was in sleepless in Seattle, I had to think about now the movie who was in it. I couldn't remember. And I mean, in retrospect, I wasn't that concerned thinking, ah, it's not that bad. And it was actually quite remarkable. This lasted about two and a half, three weeks. And now I think, I think at least according to those around me, I have gained most of my memory back. But now I have the fatigue picture. So, as much as I don't know where I picked up the virus, all three of us picked it up. And as much asI feel like I have survivor's guilt right now in the sense that, you know, I'm not that concerned about getting infected in a public exposure given I probably have some pretty good protection, at least for a few more weeks. But nonetheless, I think this potential fatigue issue is really a challenge.Eric (07:52):Yeah. The things that you're bringing up with this, like for example, I know you had had, the initial series and three boosters including the bivalent. Was that sometime in September last year, or,Michael (08:04):Yeah, it was seven and a half months before,Eric (08:07):Yeah. So,Michael (08:07):So, so that was, and I tried to get it at six months in the second. But in Minnesota we actually have a registry. And so it's not just your white card that, you know, you could do it. And it wasn't, I was trying to do something illegal, but you know, this vaccine's just sitting there. So I tried to get another bivalent at six months post my first one, and of course I was turned down. And then, five weeks after that I got covid.Eric (08:33):Yeah. And, and then of course, just recently the FDA and CDC finally came to the conclusion that for people of our age group and immunocompromised, they certainly have the option that you've advocated for. And unfortunately, you weren't able to get at that time. Although I suspect the protection, you might comment on that, Mike, that there is some protection infection for the first few months after a booster.Michael (09:00):Yeah. Yeah, absolutely. I mean, I think the studies that we've seen so far, at least, and particularly from those from other countries where they have remarkable follow up on databases, there is some initial evidence of protection in those first weeks against getting infected and even potential transmission. But that wanes unfortunately, quickly, and it's likely B-cell related immunity. And then I think as we all, at least believe the T-cell immunity, which we're still all trying to understand and characterize, probably kicks in and gives us protection against serious illness, hospitalizations and deaths. But as you and I have looked at even then, at six months out, you start to see some potential waning of that. And I think that's why we have a real challenge right now. I've said many, many times, we can't boost our way out of this pandemic. And I meant that not because some of us wouldn't be willing to get a vaccine every six months, but the vast majority of the population would not. And we've even seen here with the first bivalent booster dose, which we know has provided good protection against serial serious illness, hospitalizations, and deaths. Look at the very small proportion of the [age 65+] population that have taken that less than 40%. So it's a challenge that how do we get people to keep getting vaccinated? A lot of people say, I'm done. I'm, I'm done with it.Eric (10:22):Right, right. Unfortunately, especially those who are at high risk. It's really unfortunate. Now, one of the things you've done recently among many things, you covered the status of the pandemic today and some liabilities for the future. And you've been working on the future with the blueprint that you put together from people, experts around the world to try to map out, optimally managing this pandemic’s future, preparing for the next pandemic. Could you give us the skinny on that?Michael (10:52):Well, actually this was a report that is relabeled the Covid Wars put out by the Covid Crisis Group, which was a loose affiliation of 34 individuals who had agreed to help out developing basic materials with the hope that that would lead to a post pandemic commission, much like the commission we saw after nine And then the person that headed that up actually was the person who did head up the 9/11 commission also. And there was support from several foundations for this. When it became clear, after almost a year of trying to pull together lessons learned challenges to what we know and don't know, the US government was not gonna support, another commission either at the, in the legislative side of the government or in the executive branch. Both of them basically said, well, we're not really interested.(11:48):I think that's been a major mistake. But this report, which is now out, does address a number of the shortcomings that we have experienced with this pandemic. And again, you know, in a world where it's so partisan and everyone wants to blame someone for something, this was not meant to blame. This was meant to be what we classically call a hotwash, where we go back over an experience to learn from it. What could we have done differently? How could we have done it? What did we do right? How do we have to make sure that that's in place in the future? And so this plan is, is about that very thing. Now, at the same time I'm writing another book, much like the one I did, deadliest Enemies Our War, againsts Killer Germs in 2017, when I laid out what a pandemic might look like.(12:38):And this one is really to address what do we need to learn from this pandemic for the next one? And I go into a bit more in certain topic areas, than our report did much more in depth as it relates to vaccines, public health actions, lockdowns, all of those things. And so I hope that in a, you know, a few months that'll be available so that not only does it lay out what the challenges were, but, you know, given my public health experience of 48 years and having been through these, what do I think the lessons learned should be?A Major Prediction and Being Called IrresponsibleEric (13:17):I can't wait to read it. I mean, the roadmap, though, that you've pulled together, was really extraordinary. And of course, it addressed the things like pan-coronavirus vaccine and, and so many others that we can, pursue hopefully, and be also templates for the future. Now, I want to go back now since we recovering kind of the current future status, but back in March, 2020, you wrote that there would be, this is March, 2020, there would be 800,000 deaths in the next 18 months from Covid. Talk about an oracle, I mean, obviously no one would ever wanted that to see that, be actualized, but how did you, how did you know that, Mike? How did you know we were, we were in that, in, in store for such a dreaded outcome in an imminent period of time?Michael (14:13):Well, you know, let take a step back to December of 2019. You know, our center has a very active news team that basically covers infectious disease news from around the world. Even though it's inside of CDRAP, there's a thick wall between it and me, from an editorial standpoint, so I don't have any control over it. But they notified me that they were picking up information that last week in December, out of Wuhan about this emerging outbreak of unexplained pneumonia. And, you know, at that point we stayed on top of it. And of course, my first thought was, could this be a, a flu situation with an emerging flu pandemic, or was it just more coronavirus? You know right after 9/11, I spent three years as a special advisor to then Secretary of Health New Services, Tommy Thompson.(15:06):I split my time between the University of Minnesota and the government. And it was during that time that I actually participated actively in the first SARS outbreak that occurred, with regard to the US involvement. And then in 2012, I had been serving as an advisor to the royal family of the United Arab Emorys. And when merged first emerged on the Arabian Peninsula, I went over and worked, on that issue. And then in 2015, when MERS exploded, literally in Samsung Medical Center in South Korea, I was asked to come and I went to over to Seoul and help with that outbreak. So I had a, a pretty good feeling, I thought, for coronaviruses. And of course, influenza is something that I had been working on for 40 years. And so initially I was saying, I hold up, boy, I hope this is a coronavirus, because we know how to control that.(15:55):They're not, it's not that infectious. Even though the case fatality rates may be between 15 and 35%. Well, as you know, by the end of that first week in January, we had the data saying, yep, this was a coronavirus. But it was at that time that we had contacts in Wuhan and in Hong Kong, and we were basically getting information out. And then of course, following up with our colleagues in Singapore, the old flu network that was suggesting that this was a very different kind of coronavirus. This, there appeared to be…

Full show notes at the publisher

Our TOPPODCAST Picks

Follow Us

Stay Connected

Related Podcasts

Links

Stay Connected