Latest / AI Meets MedTech / AI Plus Biomarkers - Better Care, Fewer Resources
Transcript
- Urs Anliker: Welcome to AI Meets MedTech Podcast. My error is Jeroba. Your eye can reveal your risk of heart disease, and your wristwatch now can detect neurological conditions that your doctors cannot see. That's the reality we live in now. And my next guest is developing AI to make all of this possible. Urs Ankler is head of digital biomarkers at Holblink, a Swiss engineering company building the tools that turn raw biological data into a clinically meaningful. data sets. He was just in Barcelona at MedTech Meets AI conference last week. And I have a pleasure having him here today where I can ask him questions about the conference, about his work. Or it's so nice to have you on AI Meets MedTech. Thanks for having me on the show. Well, let's start with the conference right away. how was that? Tell me tell me a little bit about the conference, the overall feeling About the MedTech Meets AI conference. It was a very interesting conference. I met a lot of interesting people pushing AI forward in in the med tech field. And I saw also a little shift in that sense. But last year it was more experiments. And now the question is what value you you can bring additional to the product or also to the process you're working on. Yeah, 'cause you were right, I guess in the f a room full of people building AI, working with AI. What's your general sense? People are more excited, or maybe they're a little bit more cautious now, or maybe they're just overwhelmed because everyone is talking about AI nowadays. Yeah, it's not now the question of value. AI gives, and not anymore I can do everything or not nothing is working as I expected. So now we get to the real stuff and and understanding what we can do. And also what we cannot do with AI. Okay. So maybe can you share with me something that surprised you during the conference? Maybe a conversation, maybe the presentation. So it was was nice to it there were also success stories as usual, but also the not so successful stories. And it it showed you have a nice working prototype in your lab, but nobody is using it. They they had one expan example for Chinch if I pronounce it correctly. They had a nice decision support system for stopping ventilation. Nobody was using it beside the pilot users. So they f fall in love with the technology only with the pilots, but it was not generally applicable. So that was a surprising experience to hear that rollout is the difficult part of AI. Not the technology anymore. Actually, this is something I see all the time that rollouts are usually much harder. It doesn't matter if it's AI product or it could be even a physical product sometimes. But I wonder in this particular case, do you think it's a flaw of the design of somehow? So the people who designed it they didn't think through the full process, or you think it's just the hospital or clinicians, they don't see that value as you mentioned? It is more like the full process. They saw the the technical part or the the solution they have, but the integration in the full hospital process and how it will be used, not by the fans, but when you're all one and a half thousand people using it, they there are other questions coming up. It needs to make your lives eve easier, faster, and you don't want to open another tool. Yeah. Okay, but overall it was a success as you said, right? And People are excited about AI and in developing AI and talking about in a positive sense. Yeah. And there were really nice positive examples also. We had also one from Heldling where AI enabled the product performance we needed for for a field application. Or the Philips per person present nicely the how they integrated AI in their support processes to save a lot of money by preventive maintenance and beyond the right time, at the right spot. Yeah, actually I I interviewed Binu. He mentioned that a little bit before the conference. So that's that's really exciting. So let's move to the biomarkers because most people think of AI in healthcare as something that happens inside of the hospital. Sometimes people think it's inside of the diagnostic tool or maybe a robot. But in your world and in your work, you work with digital biomarkers. You need to find a specific I well, I guess the signal, health signal, let's call it that way, that can paint a picture to a clinician about what's happening with a person. So can you tell me a little bit more, what is that that signal first of all, how do you find that signal? What do you base your research on? And then what does it mean to have that clinical significance for a signal? So very big question. So clients are coming to us for developing a defined product for a defined application with defined success points. And what I presented up in the conference was a retinal health monitor. So it's for H A AMD. When your retina starts growing, you wanna have an injection for therapy stopping the growing of this. And this is a established biomarker. For a device which costs more than 10,000 US dollars and are only on doctor facilities. And the question was: can we make it for home use? And then you start, what does it mean to have it for home use? Meaning you have to reduce the device cost because you cannot give every patient at risk a 10,000 unit at home. So you have to go significantly lower. So the quality of the input signal or single processing complexity increases. So the question is then how good we need to be compared to these gold standards? Yeah. So and then you work it you forward, you you you you use much less expensive components. So we saved with a Telling how it works like 100 times of the costs. So it's a factor 100 times cheaper. But we have to compensate it on the single processing. And to get trust and understanding, we want to understand the full chain of processing. Also, what's physiologically happened. And so we start with traditional rule based single processing. And then where it gets tricky, where you cannot nicely describe it, there we put in the AI booster. And then we get to the nice performance what we are seeing now in clinical studies. So let's break it down. So for this particular retinal scanner, you're thinking of having as some a handheld device that a patient can use at home to scan their eye? that's what it is? Yeah, it's it's it's like like you put on glasses or you look for for the feldstech. so now I'm using Okay. And now AI helps you to Detect or AI analyzes the data coming from those g glasses. So so you get the the raw data, you process it several steps until you have an information on the retina thickness all over your eye. Yes. And now it's the the classification task is is does the retinal thickness change since since our last measurements? And here we put AI in. So it's it's very good in detecting structures, features in in an image to making classifications. Yeah. So do you think in the future we will move from the stationary ten thousand dollar devices that are only available in clinics to something like this? If you can build something at a fraction of the cost, but with as good as algorithm, I guess, as the one at the hosp the hospital or a doctor's office, do you think we will move towards at home based care with this speci specific Well also with all the applications, it's it's it's quite typical project for us. it's it you start with a very expensive equipment, you have a specialised facility and you want to move it to home use. You also have a lot of usability questions to address so that now in this case elderly people can handle the device. They can. That's very important, I guess, yes. I do like the idea, but I'm also now I'm thinking, what about the cost for patients versus cost for healthcare system? Do you think we will somehow shift the cost and put it on patients more in the future? Because all of those devices will be at home or wearables, et cetera, and patients will bear the cost? So n I think it now ha heavily depends on the healthcare system. yes. In in Europe it it will stay on on the insur insurance and you give impursement for for the device. And the idea is to make the overall treatment more cost effective. Yes. In in that sense, here we we have a a ch high a big challenge to finding the right spot for making the treatment. And one the traditional way is you go every month, every two months to the doctor, he will measure and and and Nine out of ten will say, Yeah, you're too early, please come back later. And then, it already too started again, so I'm a little bit too late. Yeah. My treatment. So you get better outcome for the patients and also less cost for the healthcare system in that sense. Because you have the devices and you get it to people on the doctor on the right time. Yeah. And you said that with Devices like this, you have to use the gold standard to measure it against. So, how close to the gold standard you keep your I'm gonna say it like quality of your of your devices. What is good enough for at home use, wearable use, et cetera? So w where where's that rule? How about that? Where's the rule? Yeah. So so this is always a discussion with our clients w where where you wanna go and it depends also on the risk profile or treatment profile on it. So on this device it's not so bad to Have a false alert, meaning sending the patient too early to the doctor. Besides it gives some additional cost in the treatment, but the doctor checks again and confirms now it's a time for treatment or send send the patient back. But if you are having a pe a heart peacemaker, it's a little different story. There you need to react immediately and have the high bar that it works. And there's golden standard what other devices are doing. Yeah. Have you worked with AD devices like hard monitors? Yes, yes, we also did it in the past. Yeah. We developed twenty years twenty years ago that helped when we had developed one device, not with AI at that time. Yeah. So let's go into AI now. Where do you see the most application of AI in your work nowadays? So we discussed now in the product, but it's also a big part in our development work where we support Our efficiency and also quality of our work with AI. And there are fields working well, and other fields which are challenging. Let's say it like this. Can you talk about challenges? So so it works the better when you can nicely describe what you want and you have a description of success and failure rate. And then when you have an open question. or you you didn't nicely describe it, then you get some blurry answer garbage. And in the presentation I I made the example where you want to write requirements. And if you just say, give me the requirements for an insulin pump, you could get the requirements for one or a merge of the insulin pumps on the market, but for for sure not for your insulin insulin pump. So you need to be very specific. Yeah, you have to be very specific and clear what what is a good answer and what is a bad answer. Then you get very good results. Like for top software development, unit testing, testing of the device, there you you get very good answers. And there where you want to be innovative, f do something new, by natural the model cannot do it. Because it merges what's already known in the world. Yes. So what you're saying that if you're working on something new, AI might not be as good at at it. But when you are trying to develop something based on as we know, on large sets of data. So it's much easier and you get better results. Yeah, and and that's as usual for a development what helping is doing. It's it's like eighty percent you have just to do make it good and and right. And the f say the fun part for us, the twenty percent where new the new stuff's happening. And there we need the human experts in the loop. doing it. They can get inspiration with AI, but they're irresponsible and doing the the the hard thinking, making new stuff. So good news for biomedical engineers. Yeah, they will be once Yeah. Yeah. Going back to biomarkers and how AI can help to analyze data quickly. If we are going to have all the wearables on us and we can monitor our health and our, as we said, health signals at all times. Do you think AI will be able to tell us quite early what's happening with the health and what we need to do? And do you think that may cause some sort of a issue that we will need more preventative care? Almost like too early versus what we have now? it's it's a a tricky and and and In somehow do we listen to to these devices? They were telling you in in ten years, it's it's so far away. you will have a heart attack if you not change your your relaxation, your eating habits. Yeah. Do you change based on that your lifestyle or not? And I doubt in some way that that you really a lot of people will do it. There are some people f following this advice, but not the mach maturity. Yeah. So you think we'll stick to the devices like what you mentioned? AMD. So that's the way you think we will be using more AI and biomarkers for specific diseases, right? Versus just lifestyle changes. Right. Or having support on the treatment to so to make better treatment. And and for humans or also for me, it's it's hard to have an abstract risk far away. Yeah. To follow it and and On the other side there is is is screen scancer screening which works okayish in in Europe in that sense. So you you get a request to to go to the screening. so maybe there will be some support. And the big disc discussion will come is about health status depending costs on on on yourself. Yes. Where is solidarity still working and where you say yeah it's Your fault. Yeah. And what about human in the loop situation? That's something we hear a lot right now. I hear it all the time when I talk to people. Human in a loop, human in a loop. But specifically in your field, where do you think the handoff to human should happen? And maybe if you can tell me what do you think will happen, let's in five, ten years, do you think we will remove the human involvement eventually from Reading the results of those biomarkers being tracked. Yeah, it at the end it's about trust on the AI model or on the quality on the AM model we we have and and also on the applications. And yes, human in the loop is is there where we are uncertain what the result is and we need a a second opinion, second check on it. And where the the AM model is sh is sh is sure and it will work nicely. Then we don't have the human always in the loop. And what we see is it's it is coming up with insulin pumps. You have this closed loop treatment. And there you human in the loop is basically the word the using of it. And he feels how his body reacts on the situation, what's happening with the device. And there he he looks and checks what's happening with the system. Yeah. So It's it's the the the tricky part is when the system does not what we are expecting the system is doing. And the human need to be able to react and need to be trained and understand what's happening to react correctly on it. Yeah. So maybe let's name a few examples of successful implementation of AI plus biomarkers and wearable devices or at home devices. So you already mentioned obviously the AMD. So that's something you used as an example. insulin pumps. What else? Where else do you see? Yeah, so we had an a little exotic one in that sense. it's not on the market yet. It's about measuring plot pressure for cats. So today you you're using a cuff on the tail and the cats hate it. They they basically want to kill you when you put it the cuff on on on the tail on it. And we developed for for a client a system which doesn't need the the cuff and we needed AI to get the last five percent in accuracy to get to the acceptance criteria are needed. So so this is also a nice success story on on us. Interesting. And good for cats. Good for cats, yes. Yeah, is there any other ones you can think of? Maybe you heard not maybe not in your company but maybe you heard at the conference anyone else is using AI in biomarkers space so it's it's like these control loops in in in the OR, the labs with with with oxygen saturation, heart rate. And these are the upcoming applications. There they're nice visibility study in place. And now it's a question to get it to adopted and and also to relax the stress of of the nurses, anesthesia, personal, everybody. So that they have only to react on the critical cases and not day to day. And as a as an engineer, how much time do you spend with clinical teams when you work on the product? It's it's it's two it depends on the phases. So at the beginning it's very important to understand the clinical team, how how how they're working, how the workflows look like, that we understand the situation where they are in, so that we don't build anything which is not usable for them in their day day to day day work. And then in the development phase, yes, we we recheck if if we are on the right path. And then at the end it's about validation if the clinical team is happy what we have built. And so then our client can can release it to the market later on. Mm-hmm. Makes sense. So if You would need to leave your audience with one idea about AI and healthcare. What that would be, and maybe you've done it at the conference already. So what is one thought that you'd like to share with the audience? So AI needs to have a value for the user. And not in sake for using AI to use AI. And this is a very important, and that's what also showed in the conference. We need to have value for using AI, meaning better effici efficiency, better outcome, more efficient treatment, so that everybody can benefit from it. And we ought to start. We will take a couple more time. Who do you think is responsible for that value creation? Do you think it's it's the company itself or has to be a collaboration between healthcare system? company and clinicians, what's your opinion? I know it's I know it's a big it's a big question. I understand it's case by case scenario. But if we took look looking at the bigger picture, considering everything you heard already, what do you think should happen? How can we create value? Who needs to talk first in order to see that the value is there? So it it it is a colibr collaboration, huh? And and and it's it's to see what where the future is going, understand where the benefit is and So it's a collaboration. We have the companies understanding the technology and the possibility with this technology. And we have the healthcare system with all their issues and challenges therein, which they need to be getting better and using less resources. in Europe, the healthcare professional is very rare. everybody is fighting for a for a very limited pool of talents. And so we need to reduce the burden or the the routine work has to go off, covered by AI. And the interesting, challenging part stays on the human. Yeah. Thank you so much. This makes the job more interesting, huh? That's true. That's true. I agree. Thank you so much, Urs. I really appreciate you coming to AI MitsMedTech. And I hope we'll hear more about your work and biomarkers you're working on and devices you're working on. So I'll keep following the progress. Thank you very much. It was a nice discussion. Thank you. This is AI Meets MedTech Podcast with Anna Nestorova.