Good morning and welcome to the session on AI and healthcare. I'm Jennifer Shanker, founder and editor-in-chief of The Innovator Global publication that connects business with technology. It's my great pleasure to be here for the 19th time at DLD. Thank you, Steffi, for having me back. It's my great pleasure to give a slightly longer intro to our panelists.
To my immediate left, I have Börner Keppler. Earlier today, the German Minister talked to us about the attempt to create a pool of brilliant minds in AI and machine learning, and certainly, Börner, you are one of them. You are the head of the Machine Learning and Data Analytics Lab at the FAU Erlangen-Nuremberg campus, and your work is benefiting from the funding of the German government in the high-tech agenda in Bavaria.
Next to him, we have great pleasure and honor to introduce Greg Lavender, Global CTO of Intel. Part of his work is to enable Intel's strategy across AI and confidential computing. We'll hear more about what confidential computing is in a few moments. I did want to mention that Technology Magazine named Greg as one of the top 10 global CTOs in the world. We are very pleased to have him here with us today.
If I can take just a minute to set the scene: at a time when we're navigating complex global health challenges from pandemics to chronic disease, the fusion of technological breakthroughs and rich, valuable data promises to open a vital gateway to a model of care that is as finely tuned to individual needs as the rest of our digital interaction. We've all heard of AI advances that promise to radically improve drug discovery and medical research, but to fully leverage AI and usher in personalized healthcare, we need integrated data systems because they're essential for providing a 360-degree view of patients and enabling accurate diagnosis, predicting health risk, and crafting highly tailored treatment plans.
A system that allows for seamless data sharing among healthcare stakeholders is vital, but we must also protect the privacy of patients and empower them to control their data. That is a big challenge and it involves not just technology breakthroughs but also regulatory and societal issues. It's complicated. The two people on our panel are involved in specific projects that promise to help take this forward, and we're going to hear from them about what they're doing.
Börner, let's start with you. You are working on a project with Gaia-X here in Europe on electronic health records. Tell us a little bit first about Gaia-X, what is it for people who don't know, and secondly, what are you working on and how is it going to bring things forward?
Thank you very much, Jennifer, first of all for the very nice introduction. Let me not start with Gaia-X because it's a bit controversial. I have to admit in front of everyone, so don't tell anyone, please, it's our secret here in the room, that the Gaia-X funding we took and we built something great, but it's not necessarily really supporting the future development of Gaia-X, but the future developments of something that I would call an internet of health data.
This is very important because all the big breakthroughs—I mean, DLD has been set up 20 years ago for advancing thoughts about the internet and everything that it brings to society—are tremendously important for today's AI. If you want to create foundation models, then you need the data from the internet: petabytes of text data, image data, video data, in order to create those LLMs and those foundation models. But in healthcare, we have a problem.
Let me explain that to you by a good example. Does everybody know where his or her right index finger is? Yeah, you know it, right? Greg, you know it, right? Can you give it to me? You don't want to? I need... Jennifer, do you dare? Now we could do something that we call in Bavaria 'Fingerhakeln', I won't do that with you, but it was easy, right? You chose to give me your index finger and to shake it with me.
Now if I ask any one of you, do you know where your last medical image that was taken is stored or when it was actually stored? You know it? Can you give it to me as a researcher that you trust or don't trust? Can you make an informed decision about that? It will take a long time and it's virtually very hard. I have a lot of experience with that, especially working with government organizations, hospitals, asking them for the rich data sets that are siloed inside their hospital or German health research data center environments.
So we need to get away from that for the future internet of health data. In order to create that, we have used the Gaia-X funding and a few other fundings from the biosphere, high-tech agenda initiative, and also from European projects in order to create something that inverts the current siloed data problem. Under GDPR, the right to share data is yours, every single person's right to share data with somebody you trust. But if I now have to ask for healthcare data and you say I want to give it to you or anyone else, you cannot because I have to go through the hospital, and this takes me one or one and a half years.
I've recently done that with a German hospital, takes forever and it's very hard to even get done. So in future, we want to ask you whether we can get your healthcare data for research, for AI and medicine research. This is what I'm really interested in, in personalized medicine. But in order to get this going, especially for longitudinal data sets, we need this kind of internet of health things.
With the Gaia-X funding and a few other projects, we have got this going. Now we are at the end of these funded projects. We have a working implementation that allows individuals to store their healthcare data on a personal health data cloud and then also share it under privacy-preserving considerations with researchers in Europe but also in the world.
This is a privacy-preserving implementation because we can leverage blockchain-based mechanisms. We're not implementing a blockchain, but something very similar where every interaction with the data is ledgered, meaning that it can also be approved and also be in retrospect analyzed whether the interaction with the data was correctly done, was with the intended purpose or not.
This is something I guess that will change the way that people will be able to share healthcare data, to give me or anyone else that wants to research with it your index finger, AKA this valuable information. Last but not least, I predict that the children of the future—so with Minister Blumer, maybe that's not the case yet—but they will be born with access to this personal health data space, and for them it will be as easy as giving you the index finger to give access to health data in this internet of health data. Thank you.
Thank you for that overview. Now, Greg, he used a couple of phrases that I know resonate with you because Intel is working on privacy-preserving technology and confidential computing. Tell us what those terms mean at Intel and the projects you're working on.
So we're all familiar with encrypting all of our data on disk. We use algorithms like AES to do that with larger key lengths. When we communicate over the internet through our mail, web browsers, etc., we use HTTPS, that's built on something called Transport Layer Security, TLS. That uses encryption algorithms. The whole industry is shifting to post-quantum crypto algorithms that would be resistant to quantum computer attacks.
But there's always been this sort of gap, which is what happens to your data when it's in memory before it goes into the execution unit of a CPU? How do you protect the data in memory? Because a lot of the cyber attacks that happen today are known as sophisticated side-channel attacks which attack the fundamental architecture and chip architecture of computing systems, whether it's your phone, servers, your laptops, etc.
So Intel and other companies have pioneered different approaches to this. We have something called Trusted Domain Extensions. It's in our latest Xeon Gen 5 and Gen 6 Xeon processors. We just announced earlier this week together with Google that they've launched confidential VMs and confidential containers in Google Cloud. There's a YouTube video of myself and Nelly Porter who runs confidential computing at Google. We're also working with Microsoft and other providers of this technology.
Essentially, that way we can encrypt the data while it's sitting in the dynamic random access memory of your computer, and it's only decrypted at hardware speed right before we execute an instruction to access and manipulate that data. So that's the fundamental chip architecture that happens. Now all the Linux community has been enabled with this capability. More and more applications are being built on top of it, particularly in edge computing where you're outside the normal security perimeter of your data centers, your cloud, and therefore you're more vulnerable to cyber attack or takeover of your data.
There have been examples of people spending millions of dollars to train a model deployed at the edge and it's stolen along with all the parameters. So you invested to give somebody else a good model and you want to protect that. So we get into the healthcare business, and I think Börner's closing comments were right on the money with what you just said. So how do we protect that? Well, we've actually developed an open-source library called OpenFL, Open Federated Learning. It's open-source code.
If I start at the leftmost top example, I'll give you an example of cancer research in space. We partnered together with NASA, a group called the Frontier Development Laboratory, it's kind of a leading-edge AI team that's working with NASA. We brought this together so that with the space station, they're very concerned about cosmic rays and the effects on the astronauts. Even though the space station is shielded, they're wearing their outfits except when they're in the cabin. And of course, we just had a lot of astronauts stranded for several months in space, so they got a massive dose of cosmic rays.
So we wanted to understand what's the genetic effects, how might that change their biology, etc. You want to sort of understand the absorption rate, the diffusion rate, any mutations that go, but you don't want to transmit petabytes of data back and forth to a space station. You want to do local training at the space station, think of that as the edge, and then back at NASA we have the more powerful computing capabilities to do the federated learning.
So we'll train the medical data, they take blood all the time, they're sending back their medical results, and then we want to study that over time to see how that works. But of course, it's all private information, and so the data that we use from different medical research labs has to be protected. This capability of confidential computing, whether it's virtual machines or containers, allows you to protect and preserve the privacy through physical hardware-level enforcement and encryption while it's being processed and while it's being used. And of course, it's encrypted on disk and it's encrypted in flight between the computer and the network.
So that really closes the loop. I call it the three-legged stool of confidential computing, and it's here today. We've been working on this for years. So this is sort of a landmark health study to really understand those effects.
Now if I go to this other example here, we partnered with the Perelman Medical School at the University of Pennsylvania, and they partnered with 50 institutions around the world—we said six continents, not much was going on in Antarctica around this area—and we had 71 sites. They have brain image scans looking for signs of glioblastoma tumor. I lost a family member to that cancer, so this is really important to me to get this sort of solved.
The degree of false positive rates is actually too high, but every medical institution will have its own private data about scans of brain tumors. We have data from machines from Siemens, from Philips, from General Electric, from Hitachi, and Toshiba around the world. But you can't transmit literally terabytes, petabytes of data around to do the training, so you have to do a distributed federated model.
No institution wants to share their image data, so they'll train locally and they'll generate the parameters and they'll generate the inferencing capabilities to get a better predictability because you essentially average out all the other data images that you share. You don't share the data, but you share the conclusions of analyzing the data to reduce the false positive rate. For every inference, we increase by 33% the accuracy for true positives as a result of using this sort of distributed approach, but not intermingling or giving up the privacy-preserving and protection mechanisms of the individual institution's data.
Let me just ask a question here. Federated learning isn't new, it's been around for a while. When I first learned about it, I found it very exciting. If we had a way to share data between companies that would not give away the crown jewels, but it would help advance things for society, why wouldn't we do it? How do the advances in hardware that you just described, is that going to help federated learning be used more often, more effectively? How does it take it further?
You have to have provable, it has to be provable privacy-preserving. And again, in the area of AI, all LLMs were trained mostly on public data. Their hallucinations are often for in-domain specific areas or worse because they don't have access to private data that has a higher quality, higher efficacy rate. So the challenge really then is no one wants to give up the data because the data now has massive economic value. And every enterprise organization I've ever spoken to in my career has a massive data problem. Everybody has a data lake, I call it a data swamp. They don't know where the high-value data is.
This helps you actually identify that data and put the right protections around it all the way up into the physical memory of the machine. So you got to be part of the privacy is getting provable security and provable confidentiality around it, otherwise no one trusts it.
Yeah, so this is an important element. You may have a comment on that as well.
Yeah, it's I think all about trust, trust in the implementation. From this project that I just described, it's called TeamX and HealthX, there's two consortia in Germany. Amongst the members are Charité and Fraunhofer and FAU of course, and Siemens Healthineers. We have built this secure federated learning architecture that is now in a follow-up project with the European Union. It's called DigiMed. The German Ministry of Health is amongst the scientific advisory panel members, so they're looking closely at what we're doing.
We are using this implementation to implement our federated learning architecture in various university hospitals in Europe but also in the United States. Despite the fact that we have developed this framework in our 20 million euro European project framework, and we have even found it—for all of you that are interested—a non-profit association, the European Health Data Alliance, very closely connected of course to European Health Data Space activities.
This is, I would say, trustworthy. But nevertheless, when we want to implement it in Harvard Medical School, we have to interact, or we are interacting with the data protection officers there, and it's still a matter of winning their trust, getting those things implemented, making sure that they believe that the data is never leaving the premises of Harvard Medical School, Karolinska University Hospital, or wherever else.
