The FitTech Club Podcast

"Small data, big results" - Jeff Rogers, Global Research Leader for Digital Health, IBM

• FitTech Club • Season 2 • Episode 5

Share episode

🔥 What's in this episode?

35-40% slower disease progression with early detection - this is how AI is transforming healthcare outcomes for neurological conditions. In this episode, Jeff Rogers, Global Research Leader for Digital Health at IBM, reveals how companies can train effective AI models with just 100 individuals instead of massive datasets previously required.

🎤 About our guest:

Jeff Rogers is the Global Research Leader for Digital Health at IBM and a distinguished scientist. He leads IBM's efforts in combining innovations in electronics for sensing, information technologies like artificial intelligence, and simulation tools with large computing and networking to innovate in healthcare and wellness. Under his leadership, IBM has formed a 10-year partnership with Cleveland Clinic, installing the first dedicated quantum computer for healthcare research.

🎯 What you'll learn

[01:00] How IBM defines digital health and its approach to wellness and disease prevention 

[01:53] IBM's partnership with Cleveland Clinic using quantum computing for healthcare 

[03:30] Using AI to predict and treat neurological diseases like dementia 

[05:50] The reality vs. hype of AI in healthcare applications 

[13:05] Multi-agent AI and fit-for-purpose models for fitness applications 

[19:19] New business models emerging from the separation of foundation model providers and application developers 

[21:10] The future of AI-powered chatbots for personalized health insights

📌 Links & resources

IBM Digital Health Research

Connect with Jeff Rogers

🎧 Enjoyed this episode?

Speaker 1:

Hi, I'm Natalia Karbasova, founder and CEO of the FitTech Club, a global business club for founders and executives in the fitness and consumer health tech industries. In this episode, we explore the emerging business opportunities in AI and digital health, focusing on preventive wellness. Jeff Rogers, global research leader for digital Health at IBM, reveals how IBM is pioneering fit-for-purpose AI models that create new partnership opportunities for fitness companies with significantly smaller data requirements. So, without further ado, let's dive right in.

Speaker 2:

Wonderful. We're speaking today to Jeff Rogers, who is the Global Research Leader for Digital Health at IBM, as well as Distinguished Scientist, as well as probably 10 other titles I'm not aware of. Jeff, welcome to this podcast and let's just start with you telling us what do you actually do at IBM.

Speaker 3:

Good morning, natalie. It's nice to hear from you again. At IBM, my role is to do a couple things. One is to organize the company's efforts in what we term as digital health. So that means how we bring together innovations in electronics for sensing and things like that, information technologies like artificial intelligence and other simulation tools with large computing and networking. So how do we bring together all the pieces of information technology that we have today to innovate in health care and wellness Right? So we want to think about how do you make people, how do you prevent disease from ever getting started? And when it does get started, you know, how can you use modern tools to actually help improve their treatment and help them get a better outcome?

Speaker 2:

Jeff, and what is it that you're currently working on? What's your most important project right now?

Speaker 3:

and my team have a large partnership with Cleveland Clinic now, where we've tied up over the next 10 years to take advantage of new ideas in computing like quantum computing. And Cleveland Clinic has installed one of IBM's quantum computers the first dedicated quantum computer for healthcare and they're working on their research side. We've installed that not my team, but the quantum part of IBM has installed the quantum computer there and then my team is working with researchers and doctors at Cleveland Clinic on a variety of things from how do you capture when people are having an epileptic seizure and then grade it, Because that's very hard. Today, what you ask people to do is take a pen and paper and fill out I've had a seizure and how long it lasted. But the problem with that, of course, is that a seizure, by definition, is an alteration of the mental state, so you're asking somebody to write down an experience where largely they can't. So how can we use digital health solutions to actually pick up a seizure has started, how long it lasted and what type it was? So that's an example of the things we're working on, but we're doing other things in diabetes, in Crohn's disease, in infants with malformed parts that have low cardiac output, so a range of projects we're working on with Cleveland Clinic. That's something that I find that's really exciting right now. And then the other thing that I find that's very exciting is other projects we have going on with other partners around.

Speaker 3:

You know neurological diseases. You know it's in the press a lot how dementia, various types, Alzheimer's or Lewy body and other things are actually becoming a growing problem. So how can we predict the onset of this in advance? Because there are now medications that can slow progression by 35 to 40 percent. So you know we can move neurological diseases like dementia into the chronic disease, as opposed to we're just watching you deteriorate. We're going to tell you how long it's going to take. We can now start to treat that. So for that to be successful, you have to pick up that somebody is beginning to transition and then get them on the medication and give them proper dosing, because there's bad side effects. You know, sometimes side effects aren't a big deal. These side effects are. So how can we give people the proper medication for what they're dealing with, and as early as possible? So that's a place where we're really focused.

Speaker 2:

That sounds extremely promising and exciting.

Speaker 3:

I'd still I'm trying to find the link to what you guys do in fitness and wellness. How is this connected? Short, often right, and so we can know whether or not we're successful pretty quickly. But we want to prevent disease from ever getting started. So fitness is thought to be and I say thought because it's not actually worked out is thought to really affect health, long-term health, okay, and improve it. So what we want to do is understand when a body is working well, how it is injured, when it is, how it recovers from that, and that gives us a whole different window into what it means when somebody is diseased, right. So if we understand the health perspective, that lets us look at disease in a different way and we can bring the two together so that we can actually achieve the goal of prevention of disease. So that's going to require having to do with health.

Speaker 2:

And how is AI integrated into your efforts of disease treatment and prevention? Tell us a bit more about that, please.

Speaker 3:

Yeah, so artificial intelligence one is badly overhyped. So artificial intelligence one is badly overhyped. However, what it does that's amazing is that it allows you to deal with very noisy data and to find correlations between things that you would never see another way, right? You simply couldn't do it. Now, correlation is not causation, right? Which means that just because I see a correlation between two things doesn't mean they're related, even if the correlation is strong, it can be, and often is just coincidental or somebody used the AI techniques wrong. So you know, we focus on how can we take advantage of artificial intelligence? Because it allows us to deal with noisy signals and they're always noisy from people and then also, it gives us, especially with language, it gives us new capabilities, right? So the ability to actually and this is something we're doing right now, natalie that we're excited about the National Institutes of Health in the US has funded IBM, mount Sinai and Harvard to look at kids when they come into the emergency room who were at risk for psychosis, right?

Speaker 3:

So that time frame when you're an adolescent becoming an adult is when psychosis actually gets started for a lot of people in a crisis, and we're actually using speech from them to assess are they transitioning towards schizophrenia or mania, or is this something totally else? And what's interesting is you know the way the mind generates language is a wonderful window into how it's functioning, okay. And so people who are transitioning to mania begin to use the personal pronoun on everything. So they will be, or they never use it at all, and so this is a signature that actually. And they have a flight of ideas. So they start to go through all these ideas and they make basically a flower in an embedding space, and so we can get that out of just a couple of minutes of language.

Speaker 3:

So what we're trying to do is show and the NIH in the US is funding us to demonstrate it at scale because we've done it in clinical trials small ones so now they want it at scale. So AI will give us the chance to actually understand the language they're generating right and basically say this is a person at risk for psychosis or they've already transitioned and you have to get some services behind them. So that's an example where I think you know it does something mental state and then use that evaluation of their mini mental state to actually understand is this a person who's having a psychotic event or not? And then that allows the doctors to treat them more successfully and spend a lot less time trying to get a handle on where they're going.

Speaker 2:

There is useful part of why there is some overhyped part of AI as well, and you also mentioned that medical chatbots, in your opinion, are purely overhyped. Tell us more about medical.

Speaker 3:

What I want to say is that you know, there has been great advances recently in chatbots' ways to interact with AI. Okay, and so those chatbots allow anybody to. You know, basically type in a question to an AI and get a response which is amazing. But you got to remember, it's not intelligence. Okay, no one has achieved general intelligence with AI and we're not about to Right, so claims to the contrary are marketing, in my opinion.

Speaker 3:

So the you know some very big name people are making some claims that just seem to be coming from a fantasy land Right, land right. Let's not name the names, yeah, but the reality is that there's been great advances here, but what the engines do still right, is that they put together a word and say you know, the most likely next word for a certain goal is this, and then this, or you know a sentence, you know a block of words, most likely look like this. Now, that can convince you that you know something's intelligent. Right, but you know, people should look back into the 1960s, when we first thought that we had achieved general intelligence.

Speaker 2:

OK, I have the Turing test, you have the Chinese room experiment.

Speaker 3:

It depends, and they fail that stuff right, but almost anybody doesn't actually apply the Turing test to their AI to figure out if it's right or not, if it's actually thinking right. There are some publications out now who have done it and they have shown that these ones are not thinking right. What they're doing is they're putting down random words. It's exciting what it can do, but the day-.

Speaker 2:

How do you know that you're thinking?

Speaker 3:

Well, I'm sure there's a number of people who'd say I'm not, but I think you can pass the Turing test. So it's an important question. I don't mean to be flippant about it, I don't know the answer to it, but the interesting thing is that when one of these things has what we call hallucination, you know it picks a set of words and now it goes off on a tangent. In a healthcare situation you can't have your medical chatbot doing that, right, Because if it's, you know a suggestion which seems convincing because all of them seem so convincing and it's just completely wrong. Right, it could have some really bad health effects. Right, that could kill somebody.

Speaker 3:

So that's a focus in the US, for sure, around having from our ARPA-H agency actually has had call for proposals and they're working actively funding programs to try and beat down the hallucinations. But I think the most recent advances in multi-agent AIs will allow us to actually fix that problem. So I'm hopeful that we've actually already found a solution. But what worries me is that you know people, because it's easily accessible, they convince themselves oh, I understand this, I can go and do it, Right. And what they don't quite understand is how it can hallucinate, how to really work with it properly, and what they end up is putting themselves in a situation where they're going to provide something that looks like it works but it doesn't. So you have to test stuff and you need to work with people who understand artificial intelligence, not just at the superficial level, and sometimes medical experts. Often medical experts don't, and they'll be convinced. This is amazing. I'm going to go use it with patients, right? So it needs to be tested, that's all.

Speaker 2:

So this multi-agent AI. Can you bring some very specific examples how this could be used? First of all, what is this? And second, how can this be used in the context of wellness, fitness?

Speaker 3:

consumer health. Any one AI might hallucinate, but you use multiple AIs with the same task and allow them to interact with one another and then collectively between the group, you come up with what your response is going to be Okay. And so this is a way to work around the whole idea of hallucinations, because if any one AI goes off on a hallucination, you know the other ones and you want to use different models. Ok, you want, you don't want to have the same model all the time, right, but use different models and allow them to interact. No-transcript significant right.

Speaker 3:

So if you want to field it and you don't have a big data center behind you and you don't have strong connectivity, you might want just a fit-for-purpose AI right, which is a much smaller model that has a particular goal but still gives you all the capabilities you're after. But it's just a fit for purpose one. It's smaller, easier to compute. So building those fit for purpose AIs foundation models right, fit for purpose foundation models which you can then tune, I think that's a great way. It's particularly wonderful in healthcare and fitness because you know you don't have to have a hundred thousand, a million people you train on. You train on a relatively a huge set that is not in your disease area or in your healthcare application or wellness application, and then you tailor it with a much smaller data set that is specific to your case right. So that means you know you could work with 100 individuals, where in the past that would have been a nonsensical number.

Speaker 2:

So you're speaking about fits-for-purpose AI?

Speaker 3:

Yes, and then multi-age AI. I meant to cover that in my initial comments about you know you use the multi-agents in a dialogue amongst one another. But the other direction that I think is exciting, there is around fit-for-purpose AIs.

Speaker 2:

Give me some meat. So how can, let's say, a wellness app, a fitness app that's very much on the lifestyle side of things, not that much on the disease side of things, can use these models specifically as a couple of ideas to provide better value to their customers?

Speaker 3:

Okay. So in the fit for purpose, one right in the foundation models that are built for a particular purpose. So let's say that you want to monitor someone while they're doing rehabilitation Okay. Monitor someone while they're doing rehabilitation Okay. But you and you want to use an AI to track their motion right and understand are they able to extend their arms as far as they're trying Because they injured their shoulder right, or trying to, you know, reach a certain flexibility, and so we want to see if they're doing that. So currently, an AI model you build to do that would be based on the particular cameras you're using to image and their location. You have to build it for the setup. You have to build it for the particular configuration you're using, all right.

Speaker 2:

That sounds like EGEM Genus. That's what you have to do today, right.

Speaker 3:

But with these foundation models the idea would be you actually work out a whole bunch of different configurations in the foundation model and then it's not dependent on a particular configuration. So you could allow someone, instead of having a rigid setup in a training facility, you could just set up a cell phone somewhere and actually get a valid model of them just from the cell phone right. So break the dependency on the configuration and everything by using a foundation model.

Speaker 2:

Right, I feel there's quite some things that are already happening in this space and I'm very excited to see where we land, moving forward with this kind of model?

Speaker 3:

What are those? What are in this space? What do you see?

Speaker 2:

Well, I'm not sure if this is like, what kind of model that corresponds to, right? But, already several years ago we had VEI, which was a Swiss company that exited to to nautilus, that is the manufacturer of equipment for home fitness specifically and, uh, it was about movement tracking and movement recognition for rehab, for for fitness training as well. And there are quite some other companies out there, like zing coach that also integrates ai for as computer vision on your on your phone, right so it seems that you're speaking about similar things and it's interesting to see where the differences would be, but for that you probably need to dive into the code.

Speaker 3:

No, no, no, I'm sorry. What I'm trying to say is that the performance they get using traditional methods, deep learning or any of those techniques, is inferior performance to what you would get with a model. Okay, uh, because, um, their models that they use from deep learning are dependent on the camera setup, whether they can try and mitigate that however they like, but they are. And so, uh, the foundation model would allow you to include all these different configurations in it and there's not another that you couldn't do that with a deep learning approach.

Speaker 2:

Right. So it seems like the foundation model also allows us to think about new business models as well that are connected to the technology. What's?

Speaker 3:

the most exciting business model innovation that you see in the space, moving forward with those models, wow, okay, so that's unexpected Business model innovation.

Speaker 3:

I think a separation is occurring right when you're seeing, instead of people supply an end-to-end solution, right where they give the whole integrated approach, right from the core compute through the AI model to the solution.

Speaker 3:

I think what you're seeing is a separation where some of these companies like Meta or even IBM, where I'm at you know we have the Granite models, a line of foundation models. We're providing those foundational models. Right, they're called foundation models, but we supply these big trained AI models and then people can tailor those to their particular application. So there's now a business where they don't have to go and try and compete with whoever Meta or OpenAI or IBM or anybody around training the core model. You can pick up that model and build your business on training it for your particular application, right? So I think that separation is an interesting set of business opportunities that haven't been there before. So if you're a smaller company, you can actually build on all this infrastructure without having to try and take on all the compute and all the expertise that went into that. You can tailor to your business much more.

Speaker 3:

So I think that's an interesting thing that separation.

Speaker 2:

Yeah, it's interesting and, jeff, thinking of emerging opportunities in AI on the technology side of things, on the business side of things, on the healthcare side of things as well in the next couple of years. What are you most excited about?

Speaker 3:

So I think I actually as critical as I would be of the medical chatbots, I think that's a place where we can really have innovation. Right, it will surprise people, but in the end, just like in fitness, in healthcare an awful lot of the follow-up is not quantitative, it's a survey. You give people surveys, right, and you call them on the phone and you just you ask them you know how much pain are you in, right, and it's a number. Give me a number from one to 10. And then treatment is guided by that. So I think that we're innovate.

Speaker 3:

The innovation that's coming, that we're going to take advantage of soonest, is actually going to be those chatbots interacting with people and being able to pull out insights that you simply can't get another way, and yet hold conversations with people that are empathetic, and I have watched people under treatment actually some, not everybody, but there's a handful of people or a percentage of them that will actually really enjoy talking to chatbots. It's, you know, they play with it and some people actually use it like a therapist, right, and so that's where there's going to be the earliest innovation. There's already innovation around continuous glucose monitoring, and that's for both fitness and for people with diabetes, right, and so that's going to continue. Those AIs are going to continue developing, but that's a very simple one. The one that I think will be complicated but have a very big impact is actually interacting with people and facilitating a more insightful interaction.

Speaker 3:

Right and what's the most underestimated trend that's coming up that no one is seeing yet? I don't know. I mean, people are seeing a lot of stuff. I actually do think it's related to artificial intelligence, but not in it taking over the world or doing any such thing Right, but much more around giving people a way to interact with things you know for to write, to get through their day, get through their day right, to schedule things, to understand in the fitness regime.

Speaker 3:

You know how do I get better? Or you know I have this goal. How do I get there? How do I lose weight? Right, but actually having an AI that would help you get over the little things about your behavior that prevent people from doing it right People needing immediate gratification is often what it is, and I do think that there's a chance for those assistants to be something that allows us to actually do amazing things that we haven't. That have always been a problem, like you know losing weight or taking medication regularly, or managing you know, managing blood pressure these things that would really help out. I think that that's where there's a chance.

Speaker 2:

As usual, behavior change and healthy habit formation.

Speaker 3:

Yes, there you go. That's a much better way to say it, and I think that AIs offer us a way to do that which we haven't done in the past, and also the idea along those lines of you know, a lot of people have mental health needs and there just simply is not enough of a resource to address it. I actually think AI and you know things like cognitive behavioral therapy, you know the idea that you could do that on everybody's phone and give them a way to really get more balanced. I think this is an opportunity that people aren't seeing enough.

Speaker 2:

Amazing insights. Jeff, thanks a lot.

Speaker 3:

Yeah, take care, natalie. Good to talk to you.