The AI Supercomputer Transforming Pediatric Care with Real-Time Data Access (Timothy Chou)

Full transcript

Welcome back to the podcast guys. Today we are joined by Dr. Timothy Chu, a founder of Pediatric Moonshot. Tim, welcome to the podcast. >> Glad to be here. >> To start, could you share a bit about Pediatric Moonshot and why you felt this was the project worth coming out of retirement for? >> Yeah. Uh, let me just start the story with I have no medical background whatsoever. I retired as the first president of Oracle's cloud computing business. I went back to Stanford, started a class in cloud computing and it's there maybe now 10 years ago I meet a strange student who happens to be head of pediatric cardiology at the Children's Hospital of Orange County. I like to say he adopted me. So I started learning things like, "Oh, you're kidding me. You guys are still using CDROMs," which yes they are. And then at the other end of the spectrum, and I think everybody's kind of cognizant of this today, if you're going to go build accurate AI applications, you have to have access to large quantities of data in order to train these things. And so, right around COVID time frame, I'm sitting around thinking, uh, do something useful or watch Netflix all day.

So I chose do something useful, grabbed a group of engineers and we gave ourselves a mission to reduce health care inequity, lower cost, improve outcomes for children locally, rurally and globally. How are we going to go do this? by creating real time privacy preserving AI applications based on access to data in all 1 million health care machines in all 500 children's hospitals in the world and that is what we call the pediatric moonshot. >> That's awesome. When you say reduce inequity and improve outcomes, is there any real concrete example you could give? Yeah, in the United States, 60% of the rural counties have no pediatric expertise whatsoever. Zero. Three states have no pediatric emergency physicians. 86% of the rural counties have no pediatric cardiologist. I mean, it goes even deeper. We have 3,000 pediatric cardiologists in the United States. What I just explained to you is they all live in the major metros, right?

If you leave the US, you go to India, there's 300 pediatric cardiologists. And if you go to Rwanda, there's one guy. So if you're lucky enough, the way we say it is if your mom works for Google and you live in Palo Alto, then maybe what we're working on is not that interesting. But most of the world does not profile that way. And so how do you take this very scarce resource in pediatric cancer, nefrology, oncology, all across the board here, and how do you project that into rural communities across the planet? How do you do that? It's not go build a bunch of medical schools and train a bunch of doctors and incent them to live in rural communities. The answer is let's go build AI applications that can take the knowledge expertise of these people and in essence put it in at the point of care in rural Northern California. You don't really have to go very far. I mean we've worked with a hospital in Northern Northern California in Willlets population 5,000. There's a hospital there. 20 times a month they ship a kid 4 hours away to UCSF, Stanford, Davis, etc.

because they have no pediatric expertise whatsoever. Like I said, you don't have to go very far to find how inequitable the whole system is because of where you live and the other answer is obviously what your income level is. Right. So really the data was there and you're trying to fill the gap from my understanding. >> Well let let's talk about what you can do with all of this. So let's just start with something heart disease. Congenal heart disease is actually affects one in 100 kids born in the US. Outside the US this a major cause of death. It's the number two cause of mortality in Mexico. Okay. How do you detect congenital heart disease? Well, you can because every mom goes through a fetal echo. You can get fetal echo data and you can now use fetal echo data to start to locate one of 23 different structural anomalies in the heart. This can be done. The problem is that while there are research projects, in fact, there's a team at UCSF who's done this where they have trained an AI algorithm, an AI application on a couple thousand ultrasounds and it's produced a beautiful rock curve for your data science people out there.

It's published in Nature magazine, but there it sits. It never helps any kid in rural California or Rwanda. And by the way, probably not very accurate because it was only trained on a thousand images. The question I always ask people is I go, if you're in San Francisco, you see Whimo cars driving around all the time now, autonomous driving, right? I say to people, let's assume you take that Whimo car and move it to London. Do you think it would drive very accurately? The answer is hell no. It's never known anything about driving on the other side of the road at all. So if you cannot train on large quantities of data, there's no way to go do this. And so what we're in the process of building is an AI supercomput for children's medicine. I'll just give you a simple example. The uh 500 children's hospitals in the world generate 6 million terabytes of data a year. in the cardiology department, ultrasound imaging, just for sake of reference, GPT4 was trained on a thousand terabytes of text data, right? Okay. 6 million terabytes.

Well, you're never going to move that to some giant data center in in Louisiana or I guess Elon says in space. That's not going to happen, right? The data sides are enormous. the security requirements much tougher privacy you're sitting in Europe I mean Norway is saying Norwegian data is never leaving Norway right so we came at the problem the opposite way we said well computer science 101 rather than try to move the data from to the application let's move the application to the data so we've engineered a distributed AI cloud infrastructure which literally puts cloud servers inside the building of a clinic a hospital a research lab, a home, an ambulance, which remember I said we wanted to build real time systems as real time access to ultrasound imaging. So now you can learn on 6 million terabytes of data or more. And oh, by the way, what you learn is now deployable at the point of care. So now I've built a cenal heart disease detection algorithm for fetal echoes.

I can run on every machine on the planet and do early detection of congenal heart disease. And you could do examples from many number of different areas where if you could train on large data sets, you could build very accurate applications which now could be deployed right anywhere, rural communities, large hospitals, anywhere. Do you think that this AI superco computer could be built 10 years ago or is that something that the AI space a lot? >> I think the answer to all this stuff is kind of yes and no. Yeah, anything is possible. I would say only until we started hacking at this problem we come to realize that we have to build a new infrastructure, right? A distributed infrastructure. So when I say we're building an AI supercomputer for children's medicine, what am I saying? There'll be 32 sites, 32 children's hospitals with aggregated 3,000 servers across all of them, aggregated 2,000 terabytes flowing through it, which now I can use as a supercomputer, meaning I can train a oncology application, I can train a neurology application on infinitely more data than is accessible today and deploy it to the point of care.

So, could it have been done 10 years ago? Maybe I it's kind of hard to answer that question. I think we have come to the realization that this is the approach and I don't know that anybody else has ended up there. >> Tim, you've also talked about patient digital twins and real time AI agents for a clinician or a hospital leader listening. What does a real time agent actually do at the point of care and where do you want humans firmly in control? Let me just make a broad comment about AI agents whether that's imaging or as I'm going to talk about more textoriented really I think how we're thinking about is these applications are going to end up being I'll call them red green yellow applications red it's pretty certain this condition exists this something ought to be done now green yeah everything looks okay and yellow maybe somebody really ought to take a look at this okay so as a broad comment I always point out to people, we talk about is there a Star Trek future or a Terminator future.

I I tend to want to believe in a Star Trek future. Even on the enterprise, right, there's a doc, right? So, I'm not trying to say everything gets replaced with computers, but I think humans are incredibly augmented powered by this technology. So, let me give you an example since you asked about what we call disease and trial precise AI agents. So in the world of biioarma, they are increasingly building specialized therapeutics or what FDA calls orphans. These are drugs designed for populations of 200,000 people or less. But by the way, this is a long-term trend. The work in mRNA, digital chemistry, digital biology is leading to the ability to progressively build drugs that are more and more personalized. Right? Not 200 million people, 200,000. in one day maybe 20,000. Right? So the challenge is though if you have a therapeutic and by the way 50% of the therapeutics approved in 2024 by the FDA were orphan therapeutics. So I have an orphan therapeutic for 200,000 people. How do I find them?

Right? They're very scarce. 200,000 even if you use the US population and a population of 350 million is not very many per person right so we said and we started in rare kidney diseases we went well okay there are therapeutics and I'll just give you an example for something called FSGS it's a rare kidney disease affects 200,000 people or less okay so then you go well how would I know a patient has FSGS so what we have designed are disease precise AI agents. So in the case of FSGS, this is literally 34, you could call them multimodal biomarkers, you could call them 34 questions, and I'll explain a patient digital twin in a second that are asked of a patient digital twin. If answered in the affirmative, gain confidence that this patient indeed has FSGS. So the disease precise agent just sits there and runs in the background and in essence grades all the patients in a clinic or hospital. If they grade over 90, highly likely the patient has FSGS and here are 34 reasons why. And here are the two approved therapeutics. If you grade under 50, you go, okay, everything's cool.

If you grade between 50 and 90, you could actually rank order the blood test, urinalysis, biopsy by cost to either gain or not confidence that this patient indeed has this condition. I said these the agent is running against a patient digital twin. So what's a patient digital twin? Well, it's in essence, you know, Zach, Tim, Mary, Bill all will have a patient digital twin. At first we are training on the electronic medical record which is the clinical view of a patient. Uh we started this work in oncology just for you to think about an oncology patient can have a medical record that is 3500 to 7,000 pages of PDF. Just for reference a box of printer paper is 5,000 pages. So there's a lot of data there. So what are we doing? We're using that data to train a local large language model privacy preserving. Now you can ask anything about that patient. We actually demonstrated this to 12 oncology nurses around the country.

We were totally surprised. I mean, as an example, a chemotherapy nurse will spend 15 to 20 minutes trying to calculate how much chemo, how much cis platin was the patient given in the last three times they were there. It takes them a long time to do that because they're literally looking at a PDF document with a search bar. Here, you just ask it. You say, "Well, how much was it given?" And it replies. You can actually talk to it, >> right? We turned this over the controls over to the nurses and they started asking questions and one of them asked the question well what is the prognosis for this patient and I okay time out that's not what it is it knows everything about the patient so we call the user interface total recall because it knows everything about the past of the patient I made the comment this is trained on the electronic medical record but remember there's a ton of data that happens outside the clinic your Fitbit, your Apple Watch, wearables, etc., all have data.

And we now have the capability to put a server inside your house, which now can take data coming from your wearable world blended with data coming from the clinical world and create a more complete version, a digital version of that person. Or as Ronnie Cohen, who's CEO at Sick Kids, we did a podcast with him a couple months ago. He said, "Their mission is to create precision health care for kids by using data from the zip code to the genetic code." And I said, 'Yeah, that if you had zip code to genetic code data about Zach, Tim, Mary, Bill, there's all sorts of things you can do, including what we just said, be able to diagnose a whole series of diseases using computers, not having to depend on humans understanding all of this. By the way, the FSGS lecture in med school is one whole hour. So, we talk about this. I don't know how many your listeners have ever watched House, but we talk about putting Dr. House in the house. Literally, you can do this across 6,500 rare diseases, right? And then the clinical trial problem.

I can now select people automatically and make them eligible for clinical trials because you have the complete knowledge of this patient. Where are they located? What's their zip code? What's their genetic code? What's their blood pressure? What how much did they run yesterday? And now we can do this at scale. And today, I don't know how many of your listeners know this, a lot of this happens very manually. Well, very interesting. You've launched a major capital campaign to complete the next phase over the next year. Are there any proof points that you are most focused on? and what's the best way for hospitals and researchers and supporters to get involved? >> As you pointed out, late last year, we came to the conclusion that in pediatric medicine, the only way this all gets funded is philanthropically. There's not a commercial angle to all this. And so, we launched a capital campaign.

We created Pediatric Moonshot as a 501c3 as a nonprofit. We're targeting over a hundred different foundations who have had a history of donating money to children's healthcare to complete building what we're calling an AI supercomputer for children's medicine. 32 sites, 3,000 servers, 2,000 terabytes of data flowing through it to be able to push the state-of-the-art of being able to do accelerate diagnosis and treatment in cardiology, oncology, nephrology all across the spectrum. any people who are interested in contributing, any of the foundations interested in learning more about what we're doing, we're eager to talk to them. You know, you can reach out to me there. I'm on LinkedIn, etc. You can reach out through www.pediatricshot.org, which is the website for the foundation. As far as the clinicians, etc., who might be interested in what we're working on, you can reach out again directly. There's so many places to learn more about what we're doing. I'm sure you can put it in the podcast. Pediatric moonshot.org has a subscription to a newsletter. So, you can subscribe to a newsletter.

LinkedIn, we are actually there are over 2,000 followers of the LinkedIn page that we publish into about updates and what's going on with Pediatric Moonshot. There is the pediatric moonshot YouTube channel that we're just launching. There is the pediatric moonshot podcast series available on Spotify and on Apple Music. I was just telling you we're over 40 of these right now. And we have everybody from leading clinicians, CEOs of I just mentioned Sick Kids in Toronto, Children's Mercy in Kansas City, Raidies Children's in San Diego on the podcast series. Lots of ways to learn more and by following, subscribing, etc. supporting our mission. Great, Tim. I will add links so people can just click in the description.