Breaking the Data Stratosphere
This week I am talking to Timothy Chou, PhD Board Member, Stanford Cloud Computing Lecturer and Founder of Bevel Cloud, a Pediatric Moonshot to reduce healthcare inequity, lower healthcare costs and improve outcomes for children – locally and globally by creating privacy-preserving real-time applications based on access to data in all 1,000,000 healthcare machines in all 500 children’s hospitals in the world
Timothy shared his unique journey, transitioning from an IT career, including running Oracle’s cloud computing business, to his current role in healthcare innovation. He shared his mission to tackle healthcare inequity, reduce costs, and enhance patient outcomes through Bevel Cloud’s Moonshot program to share data with pediatric-focused initiatives.
We discuss the challenges plaguing healthcare systems, such as the outdated practice of using CD-ROMs to share image data. With the space focus it was inevitable that their first two projects were called Mercury and Gemini. Mercury focuses on simplifying image sharing for pediatric cases, streamlining the process for emergency departments and clinics allowing images and data to be shared as easily as we share images on Instagram and TikTok.
Gemini involves creating a decentralized AI research lab for children’s medicine, harnessing federated learning to build accurate AI algorithms while maintaining data privacy and respecting differing data regulations around the world.
Listen in to hear how this vision extends beyond pediatric medicine, recognizing the potential for his solutions to address broader healthcare challenges.
Listen live at 4:00 AM, 12:00 Noon, or 8:00 PM ET, Monday through Friday for the next week at HealthcareNOW Radio. After that, you can listen on demand (See podcast information below.) Join the conversation on Twitter at #TheIncrementalist.
Listen along on HealthcareNowRadio or on SoundCloud
Raw Transcript
Nick van Terheyden
And today, I’m delighted to be joined by Timothy Chou. He is a board member, Stanford cloud computing lecturer, and founder of bevel cloud, a pediatric moonshot, as we do each and every week. Tim, Timothy, if you would tell us a little bit about your story to this point in your career, because it bears a lot of relevance to some of the topic that we’re going to cover today.
Timothy Chou
Yeah, well, first of all, thank you for having me. So, as I say, to my clinician, friends, I have no medical background whatsoever. I have been in the IT side of the world. Since the beginning of my career, my last job was running the cloud computing business at Oracle. When I retired, I went back to Stanford and started up the first class on cloud computing there. And that’s really the origin of all of what has gone on, or I call it my last great project. Because I meet a guy who has an MD and mph and MBA. He’s chief of pediatric cardiology at the Children’s Hospital of Orange County. And as you know, Dr. Anthony Chang has gone on to be a major, let’s call it advocate for AI in medicine, whether it’s pediatric or adult medicine, and it’s really Anthony, who is a portal into the world of medicine in the world of pediatric medicine. And I learned things go on.
Nick van Terheyden
No, I was just gonna say I would say he’s more than an advocate. He’s a force for yes. And we’re very lucky to both know Him and be around him, because it’s always a great pleasure to, to see and hear what he’s managed to achieve. When I think in many years gone by people were sort of Yeah, it’s interesting, but not that interesting. So you know, highly relevant background. And, you know, lots of experience, but you’ve really sort of narrowed this down into a focus on Bevel cloud. And if you would, tell us a little bit about what the challenge is, what the problem is that you’re solving and a little bit of what the story is today.
Timothy Chou
Yeah. So after after meeting Anthony, right, I, I came to learn that the healthcare system still uses CD ROMs to pass data around which I was like, Are you kidding me. And then, obviously, at the other end of the spectrum, if you’re student of what is going on in the world of AI, and particularly deep learning, the only way we’re ever going to build image based AI is we have to get access to large quantities of diverse data. Otherwise, you can’t do this, right. And so I’m listening to all this and my whole backgrounds, infrastructure software, and I’m thinking this is all the same problem. And so, right at the beginning of COVID, mostly because we all got bored of watching too much Netflix. I brought a team together of infrastructure, engineers, people I’ve worked with across the board. And we set ourselves a mission, we call it the pediatric moonshot, which is to reduce healthcare inequity, lower cost, and improve patient outcomes for kids locally, nationally, and globally. How, by creating privacy preserving real time applications, based on access to data in all 1 million health care machines, in all 500 children’s hospitals in the world. And when I say healthcare machine, I mean everything from imaging machines, ultrasounds, MRIs, CTS, etc. But all the way to bedside monitors, blood analyzers, Gene sequencers, all of that fits into machines, because that’s really where the data is. I mean, the EMR EHR thing is just a giant billing engine with some notes in it, which is fine. I know why we have to have that. But that’s not where the real that’s not the treasure trove. The Treasure Trove is sitting in the machines.
Nick van Terheyden
So I can’t disagree with you. I think that’s exactly right. We’ve we’ve struggled with this. And, you know, through the course of my history, I’ve lived this experience of you know, prior to the digital, you know, when we move from imaging into the digital A world that was significantly painful, we had to sort of try and grab some of the old images, scan them. You know, I’ve tried to carry this around to this day, I still receive a CD ROM from every imaging study that I get. And I imported and carry that through. Why are we in this circumstance? I mean, it just doesn’t seem like it should be that difficult yet. It clearly must be because otherwise it would have been fixed by now.
Timothy Chou
Yeah, I’ll tell you, I think some of the difficulty and I’ll talk a little bit about our technology, some of it is, we haven’t had the right technological approach until now. And then some of it, as all of you in healthcare, well recognize has to do with all the silos that are organized within hospitals across hospitals that make any ability to build networks really difficult. So it’s both sides, right. And we could get into privacy and other issues. But But let me just say, we took this mission and said, Well, how the hell are we going to do this? Right, all million machines. So just like the original moon shot, we said, we need to build a new rocket. So we built a new rocket, it is a highly decentralized in the building. Cloud Service, meaning that the servers so if you took my class on the very first day at Stanford, I would say, Well, what is Amazon or Google Cloud? It’s they buy a bunch of machines, they manage them, they put them in about 10 data centers in the world. What is bevel cloud doing? Getting a bunch of machines, managing them, that just figured out how we could put it inside the building at the Children’s Hospital of Orange County, or at Bambino J Su and the Vatican. Why does it need to be in the building, you might ask? Well, the only way you can talk to the ultrasound machine, is you have to be on the network with the ultrasound machine, or the blood analyzer, or any of the machines I talked about. So obviously, day one, we knew that security and privacy were going to be the biggest issues. And the very first conversations we had with the IT departments, they all said and FW. Yeah, right?
Yeah, we were gonna say that was gonna be Yeah,
Timothy Chou
we all Yeah, cuz we’ve all been in this world. And so we know that that’s what they’re gonna say. So we engineered over 30 different security and privacy features from the very getgo. Because we knew this so that we could securely do this, meaning be able to place a server inside the Children’s Hospital of Orange County, or, you know, any children’s hospital, or, frankly, any building, as you will hear in a minute. So 2022, we really dedicated to, okay, let’s get out of PowerPoint, and do this. And so 2022 was very much focused around deploying what we call edge zones, which is to put a server inside the Children’s Hospital in the building, and six, excuse me, eight different buildings on two, three different continents. Right. So I’ll say we exited 2022 feeling pretty good about, we knew how it could work, we had gone through security reviews, you can easily imagine this, right? Ba agreements, I mean, all of the you know, things that are necessary to put something like this in place. Now, just to make an economic point about this, because we do all of this work, security work, right? Data Use Agreement, blah, blah, blah. Now an application, a cardiology application, or radiology application does not have to go through all of that, because we’ve done the underlying infrastructure. And unfortunately, what we’ve observed up until now is that a person who goes in and wants to build a sepsis prediction algorithm ends up having to build the full stack ends up having to deal with the question of, well, what are you securely attached to the bedside monitor? And how are you going to send the data and is it encrypted with a bunch of IT questions, which obviously, we have perfected, so that the application people do not have to worry about this? Back to my analog of what are we really doing? It’s not a whole lot different than what Apple did with the iPhone, right, which is they built an infrastructure, a phone, they distributed around the world that gave you an operating environment, right to write applications with They gave you access to a camera that gave you access to the temperature that gave you access to location. And they invited the world to build applications. So we see now 10s of 1000s of applications on the phone. Fundamentally, we’re not doing anything different. We have an infrastructure, we’re deploying it globally. It’s just that our camera is an ultrasound, our cameras, an MRI, our cameras, blood analyzer. And so that’s why, beginning this year, we said, Okay, if the phone had just arrived with it, no apps on it, you and I would have gone. Oh, that’s cool. But so what, right? So we focused ourselves on two major programs, one we called mercury, and the other is Gemini. So let me describe. So the Mercury program is a byproduct of us spending a lot of time with people in emergency medicine. Now, you can guess very early on, we spent a lot of time with people in cardiology, because Anthony’s cardiologist, right. But when we spent time with emergency medicine at Texas Children’s or at UCSF, or at Nationwide, we started hearing the same thing, which is that the kids arrive in the emergency room with no image. If they do arrive, it’s on a CD ROM that they cannot read. And then the guys at Texas Children’s said, you know, if we could have seen the image ahead of time, we have to tell them, Don’t send the kid. So you’re sitting there going, Oh, come on this problem. You and I are just talking about it, probably How come it hasn’t been solved, right? So I sat down with the head of trauma at UCSF, but I said, Chris, there’s software out there. Umbra health, life image isn’t as a solved problem. You guys know, because all that stuff is too complicated. We have to send training teams to Willits, California, which is in Northern California, to train them on how to use the software. And we’re sitting around going, that makes no sense. Like, why is it that hard? So we have engineered an application, which looks remarkably like Instagram. So the tech just make a point of it. We engineered this not for like sophisticated Children’s Hospital. It’s like a clinic in a shopping center in Dallas, right? So okay, the tech box over to the X ray machine, there’s QR code, you scan it, up pops the X ray images from the past 24 hours, Select Select Select, share with Texas Children’s share with children’s of Dallas done not not much different than how everybody is using Instagram to share right images. Of course, our images are X ray images. So Mercury is all about building out a nationwide network. We’re actually in conversations with a couple of people about making this global. But it’s basically the same idea how can we simplify image sharing.
Nick van Terheyden
So for those of you just joining, I’m Dr. Nick the incrementalist today I’m talking to Timothy Chou. He’s a board member of Stanford cloud computing lecturer, and founder of bevel cloud, a pediatric moonshot. We were just hearing some of the details and specifically around Mercury, you know, I’m a space nut. So I’m obviously a big fan of this, you know, naming convention and wondering what Apollo is. But we just covered mercury. And you talked about that going global, with some, you know, opportunity in the future. But before we dive in a little bit more, tell us about Gemini.
Timothy Chou
Very good. So Gemini is the other end of the problem. So if you think about what we’re doing with mercury is it’s a very simple application that now deploys not only in children’s hospitals, but to regional hospitals, clinics, etc. Right? So we get the infrastructure in place. So now I can share with a human right to get a human’s expertise. But as we’ve all talked about, hey, there’s not enough pediatric cardiologist in the world. There’s only 300 of them in India. Rwanda has a single pediatric cardiologist. Yes, Toronto has 50 of them, but go to Thunder Bay and there’s zero. Even in our great state of California, you know, find a pediatric cardiologist in Salinas or Modesto is not so simple to do, right. So, obviously, it would make sense if we could take the knowledge of our best pediatric cardiologists are radiologists and build AI applications to go do this. But the key to building AI applications is you must have aI access to large quantities of diverse data. Up until now, even there was just a paper released by Matt Lundgren, who used to run the AI medicine program at Stanford and pronase is over at, at Harvard, and basically is a study of the state of the art, radiological AI. And it’s pretty poor, meaning they say, Hey, it works great on the test data set, but just go put it into the real world and it fails. Well, that’s because it’s no different than if I trained a car to drive in Palo Alto autonomously. And then I moved to Atlanta, and wonder why it’s crashing into things. It’s the same thing. You have not shown it enough data to train on it. But we’re sitting here going, Hey, man, there’s plenty of data, the 500 children’s hospitals in the world generate about 6 million terabytes of data a year. To put that in context. NIH has been on a five year program to build an imaging data commons. They’re at like 44 terabytes. I mean, the amount of data is are Ghannouchi one. Okay, so what’s the problem? What if we got plenty today? What’s the problem? And the problem is the approach that we’ve used to build chat GPT, or image net, or any of the things in the consumer side will not work in medicine. So the technique today is to deliver a centralized architecture. So I will bring all the images of dogs and cats, I’ll bring all the text right to a central location. And then I’ll split it into training datasets and test datasets. And I’ll train up my chat GPT. Right on that. That’s the methodology. The challenge of that is, okay, let’s go play the game of our pediatric cardiology example. Okay, 6 million terabytes. Where are we going to put it? Let’s go put it in Ireland, let’s get to a good place. Okay, who’s gonna pay data transfer? From Rome to Ireland, from Ireland, to San Francisco, from Ireland to Boston? excetera? I mean, who’s going to do that? How would you ever build a real time system sitting in Ireland? Right? And finally, what about privacy? One of our key core team members is a former student who has 15 years of privacy law experience. So we engineered around privacy management on day one, one of the central tenants of privacy management is purpose limitation. When you aggregate a whole bunch of data at some site, you gotta go, what are you gonna do with it? Ah, we’re gonna do good stuff with it. net, net, net, net net. And obviously, our European friends can talk about data sovereignty issues, data, residency issues, etc. So the trick question becomes, I got those 6 million terabytes. But how do I learn in a decentralized architecture. And so we are borrowing technology that was developed on consumer side, referred to as federated learning. Right, which powers Siri is not sending your Voiceprint to the Apple cloud. at all, it’s learning on your phone learning locally and only transmitting model parameters, right? We’ve done early work that says the same thing can be true. So we can learn on the data in the children’s hospital in Delaware without ever leaving Delaware, or in the Children’s Hospital in Chicago without ever leaving Chicago. And that is what Gemini is all about. We are building an AI Lab, a research lab for children’s medicine, which will initially start with 100, ultrasounds, twin to 100, servers on six in six locations on two continents. And we will provide a very simple neural network application referred to as ejection fraction measurement. And we will use that to help ourselves understand one watt federated learning frameworks work the best most of these were developed for the consumer side, as well as what methodologies work the best to be able to build accurate AI algorithms. And that is that we’re going to expand it to 32 sites, all imaging machines in the first year and a half of the Gemini program. So Gemini is a research lab. For us, the collective us, the clinicians, the tech people to really push the state of the art of AI and and children’s medicine. And that’s the purpose of Gemini
Nick van Terheyden
reminds me a lot of the SETI originals SETI program that distributed little modules. I know, it’s not exactly the same, but the same kind of principle of using that sort of diverse computing, and powering all of this, I mean, similar principles, obviously. Or at least, you know, to my understanding that solves all resolves some of the privacy problems, because you’re not transporting or transmitting that data, you’re allowing access or giving access through the secure channels that you’ve created through the Mercury program. You’ve repeatedly said throughout all of this paediatric? And I’m going to ask because, you know, I’m a little bit biased, you know, maybe, you know, somewhat interested, because I’m no longer in that zone. Is, is this just for pediatric? Or are we seeing we’re going to see this in, you know, broader categories?
Timothy Chou
Yeah, I have obviously been asked that question. We, you know, we’re a young venture. And one of the things about young ventures is you have to have focus. And so we feel like this is a good focus, it’s a mission worth accomplishing. But obviously, the ultrasound does not know what hospital it’s in. And I just explained that mercury is going to put edge zones in non children’s hospitals. So the techniques that we are, are implementing, are all equally applicable to adult medicine. But I, you know, back to focus, you know, let’s let’s accomplish this mission. And then, you know, it’s that the portability and adult is there,
Nick van Terheyden
fully respect that, I think, you know, there are many startups that fail because of diversification and lack of focus. So I fully understand that. So as you think about the future, and I’m gonna guess now, the next program, Apollo, what’s, what’s coming there? And you could tell me if that’s not the right now,
Timothy Chou
it is a good thing about using this convention, it lets at least for all of us all people that remember all these names, right? Oh, come on,
Nick van Terheyden
it inspires. Let’s be clear, at least it did me.
Timothy Chou
Hey, I actually declined a job at IBM to go work for NASA.
Nick van Terheyden
Oh, I have. Okay, make me even more envious.
Timothy Chou
Yeah, yeah. So you know, I wanted to be a rocket scientist. So Apollo, least for us, and we’re not putting much effort into this right now is really when you put the two things together, because what I described as a research lab, which will hopefully build a lot of very interesting applications in cardiology, radiology, etc. There’s the next step, as everybody knows, which is okay, I’ve got the software. Now, how does this go into production? Right. And really, Apollo’s where the two things come together, right, we’ve got the applications. And now we’ve got the infrastructure to layer it into, so that we can now bring the two pieces together, and we’ll be on the moon, as
Nick van Terheyden
well. And then we’ll be going further than that, because obviously, there’s many, many more applications. I mean, you’ve, you’ve taken one of those wicked problems in healthcare, which I think to those outside seems always a little bit obscure as to why that would be a wicked problem. Gosh, it’s just an image, you know, and we do we all use Instagram, and, you know, share photos, and whatever. And, you know, security plays some degree to that, but it’s not always centered on that there’s many, many layers to this, I think, you know, the thoughtful approach that you’ve identified, built for this, and created this infrastructure around a very focused problem that, you know, people can get behind. I think that’s also elegant in the, you know, pediatric sharing of images. And, you know, children are future and, you know, we’re under serving them, as you described with people arriving in emergency rooms that shouldn’t be and, you know, lack of sharing of imaging. Very exciting time. I’m obviously excited about it from a personal standpoint, but, you know, recognize it’s not going to hit my area of utility for a little bit of time. But I think the AI learning opportunity that you described around Gemini is really going to be extraordinary in terms of the opportunity because as we now centralize that data, we’ve got a much more robust data set, and importantly, deals with some of the bias that we’ve seen as a result of these very segmented or narrow focus data points. Unfortunately, as we do each and every week, we’ve just run out of time, so just remain Nice to meet you. Thank you for joining me on the show. Timothy, thanks for joining me.
Timothy Chou
Well, I really appreciate you having me and you know, getting the story out to people and we always tell people, you know, it took 40,000 people to get to the moon. So we’re always eager to have people join the mission because it isn’t going to be done which is a couple people. So thank you for being part of that.
