Latest / AI Meets MedTech / How AI Is Revolutionizing Cardiopulmonary Diagnostics with Expert-Level Precision
Transcript
- Anna Nesterova: Welcome to AI Meets MedTech. today my guest is Orr Imbar, the co-founder and CEO of Metabit, an AI-driven platform that is changing the way we're diagnose cardiopulmonary conditions. Orr, thank you so much for being here. Welcome to AI Meets MedTech. Or Inbar: Thank Anna Nesterova: Or, if I would had no idea what you do, what your company does, how would you describe your technology in a relatively short period of time? Or Inbar: So Medivite develops AI-powered software that tells physicians in the interpretation of complex cardiopulmonary tests, starting with cardiopulmonary exit testing, which is CPET, which is abbreviation. What is unique about CPET is that it's one of the most informative, rich tests in medicine. It can reveal whether a patient's symptoms are driven by the heart, the lungs, muscles, or even deconditioning. Anna Nesterova: Mm. Or Inbar: The challenge is that interpreting CPET requires years of specialized training and expertise, which is why its adoption has remained limited despite its clinical value. Many people think this is simple, an AI problem, but in reality, it's equally a physiology problem. Building an algorithm that can interpret CPET requires a deep, deep understanding of excess physiology and decades of clinical expertise. That is one of our key advantages at Medibite. We combine advanced AI with the knowledge of world-leading CPET experts, including my father, Dr. Woon-Rin, who has spent decades advancing the field. Anna Nesterova: Mm. Or Inbar: Our goal is not just to automate interpretation, but to capture expert level reasoning and make it valuable to more healthcare providers. By doing so, we can help clinicians perform more tests, save valuable physician time, and bring advanced cardiopulmonary diagnostics to many more patients. Anna Nesterova: the CPAT interpretation, as you mentioned, traditionally is very complex. And I do understand why you decided to use AI to help out with interpretation. But I wonder what was the hardest part of teaching AI to interpret something like that? Because if it's hard for a human, is it that much easier for AI to interpret, or it still takes a lot of human power to train AI? Or Inbar: specifically in CPET, in order to train the model to do the interpretation, you need first of all, a lot of preprocessing of the data which come from know-how and expertise of CPET. And in the other end, it's super machine learning. So you need... an expert label of the data, which is our advantage or also our complexity in doing that. Anna Nesterova: Mm-hmm. Or Inbar: is that receiving label data from world expert. But in the other end, also the companies that manufacture CPEC, they are not on the data. The data is owned by the hospitals, the research centers. So it's critical the label data. And we work with really the world expert in that area, which... Anna Nesterova: Mm-hmm. Yeah. Or Inbar: help us to bring the accurate model in the world. Anna Nesterova: And was it the hardest to find that data and get it from multiple resources? Was that hard or it was relatively accessible? Because I would imagine I would imagine it's a hard one because if it's a tough test to do, tough to interpret it, and then maybe different countries, different places, different clinicians maybe they have their own way of doing so. So what was the hardest in that process? Or Inbar: Very. Exactly. First of all, we knew that it will be common, but I didn't know how much. Even without names, even in the largest facilities, the world-known facilities in CPEC. The data is separated in different machines, different databases. So if you want the raw data of the CPET, you need to go to the machine. If you want the spirometry, you need to go to different machines. There is the EMR that you can need to take some data from there. It's a lot of time to arrange the data as we want it. But now with the CPET-wise, which is the platform of the CPET, all the data from now on, in facilities that work with our database will arrange in the way that it's valuable. You mean you can improve all the time the algorithm and also unutilized data, you know? Yeah. Anna Nesterova: Mm-hmm. Yeah. I have a question for you in terms of your platform. Was it C Wise, right? Is it considered to be a medical device or is it considered to be a diagnostic device? How does it work when you apply for of for regulatory approvals? Or Inbar: So from a regulatory pool, it's SAMD, software as a medical device. We were actually on the edge of a SAMD because we can influence on that. We assist the physician. We are not telling me what to decide in the end, but. Anna Nesterova: As a medical device, okay. Or Inbar: We assist him and Paul influences on his decision. I will note that our submission was through a third party because it's a low risk product. So there is a possibility to go through via third party in the submission. Anna Nesterova: Mm-hmm. Yeah. you said that the system is assisting a physician. Can you tell me more? So now right now it takes about an hour to interpret a CPAT. And it takes one, maybe two, clinicians. So what are you doing? What's your solution for that? Or Inbar: So first of all, it's the workflow that we built in order to get to the final decision. And of course, the report is automatic. this is the first thing that we did in the platform. The second is currently the software that in the market, which are a... Anna Nesterova: Uh-huh. Or Inbar: developed by the manufacturer of the metabolic code, our analysis, its graphs, raw data, and but you see in order to read this analysis, you need a lot of years of expertise. We actually, in addition to the AI and the interpretation, we bring analysis that we actually explain the physician, what is the right interpretation, the right analysis. Everything is color coding. We take him through a workflow that explains all the way the results. Anna Nesterova: Yeah, and it takes them now a few minutes. If they learn everything and then learn the platform, instead of an hour or maybe a few hours, it takes down to you think five, ten minutes. Yeah. Okay. So would you trust your own AI system to read your own test results? So there's no doctor. We we're talking about in general, in theory. If there's no doctor, no second opinion. Or Inbar: Yeah, exactly. Anna Nesterova: Would you trust the AI that you're building to read your results and interpret them? Or Inbar: Yes, I will trust and I must say that how we build the study, the validation, it will be in different levels. First, we will approve that we can separate between normal and abnormal, which is very important. Then we'll go down to the limiting system. What is the reason that the subject stopped the exercise? Anna Nesterova: Okay. Mm. Mm-hmm. Mm-hmm. Or Inbar: and then we will go to the lower level of severity of the limiting system. And after that, we will decide the detailed interpretation, which is referring its suggestion that we give, referring to a disease. So in that way, the mistake that we can do is very minimal. Anna Nesterova: Do you find that doctors find it becoming skeptical about it or maybe they're being threatened by this technology? Because part of what I do in my podcast, I'm trying to explore the future and where the industry is going with AI tools. But of course, it comes with great positives and also some negatives and some pushbacks. So I wonder if doctors you're working with they feel threatened by it or they think it's actually make is making their life easier. Or Inbar: Absolutely, 100 % is there, make their life easier. We actually take all the manual work to the side and let them focus on the interpretation, on the treatment they want to give for the patient. So... Anna Nesterova: Is here. Mm. Or Inbar: in 100 % we make their life a lot easier and also save a lot of time. Anna Nesterova: Mm-hmm. do you find any surprises while you're developing this AI tool? Anything that surprised you about AI and what happened along the way of development of this product? Or Inbar: Yes, all the time. There is a lot of challenges. As I told you before, there is a lot of preprocessing of the raw data and we have a few options for preprocessing. Our goal is to find the algorithm that only uses the raw data, okay? And not do any analysis before it's put in the model. But... Anna Nesterova: Mm-hmm. Or Inbar: there is a lot of consideration related to normal values, the reference values of the patient. So this is a lot of challenge, a lot of development that needs to be done. This is why I start with that you need both the know-how of CPET and the technical expertise of machinery. Anna Nesterova: Yeah. f let's move to the future. In your ideal world, let's say five years from now, or maybe ten years from now, I actually think now we're moving at such speed that maybe ten years it's too far away. in five seven years from now, what would be the ideal diagnostic process for people with cardiopulmonary conditions? Or Inbar: So ideally the patient perform a test and within minutes the physician has a complete analysis that is integrated with the patient medical history, imaging, lab result and previous test. Instead of spending time collecting and interpreting data, the physician can focus on understanding the patient and choosing the best treatment strategy. Anna Nesterova: Yeah. Or Inbar: The result is faster diagnosis, earlier intervention and hopefully better outcomes. This is how I see the future. Anna Nesterova: Yeah. And I would also and I would also imagine the overall amount of patients could be also increased. So more patients could be helped because you save so much time. Or Inbar: I will add to that that in the future, we also look on the wellness area and also to the sport medicine. This is an area that these tests can be done by a healthy subject and give them a lot of information of how to train, what to eat, a lot of things that these tests can bring because it's the only test that in one test can look on the three main physiological systems. Anna Nesterova: Athletics, yeah. Do you think the system will have like you mentioned that there's a different region I read on your website? You're trying to reach more specialists, more population, bigger population. How would that look like? Do you think more tests will be available in different geographical areas or more tests will be taken in those specialized facilities where trained people or maybe speci specific equipment is already placed? So how would that look like? Or Inbar: I will give you an example. Okay, we walk without again without names. We work with a former president of one of the biggest company, the manufacturer. When they, when he work in the company, they wanted some facility, small facility, wanted to do the sieve for the people in the area. Anna Nesterova: Mm-hmm. Or Inbar: They want to buy the machine, but they didn't have the expert to do the interpretation. Okay. So what we bring in addition to the times, saving and everything, we bring remote interpretation, which means that any small facility in the U.S. for example, can buy the machine, can do the test with a technician nurse and in a minute send it to an expert in the other side of the U.S. and give the community the option to receive the information and interpretation of See you then. Anna Nesterova: Yeah. No, that's fantastic. And I also found very interesting that unification of data is another next step because like we were saying before, different centers they had their own data, they had their own methods. But now with this type of system with help of AI, we can unify the data and everyone can follow the same protocol, which will allow to do the research better, to interpret results in the future better. So it seems like we are moving in the right direction, like you said, hopefully in a few years. Or Inbar: Thank Anna Nesterova: You just go in, do the test, and right away you know the result. Or Inbar: Exactly. Anna Nesterova: is there anything in healthcare currently that surprised you? Something recently that maybe you read or saw, something that fascinated you? Or Inbar: It's a big question, but what surprised me if I can put a point on one thing is what surprised me is how quickly attitudes have changed in this area. A few years ago, conversations were about whether AI could be trusted in health care. Today, many clinicians are already using AI tools in their daily work. What I have learned in that... Anna Nesterova: Mm-hmm. Okay. Or Inbar: that doctors don't want AI to replace them. They want AI to help them. And this is exactly what we are doing. When the technology is transparent and actually solves a real problem, adoption happens much faster than people expect. Anna Nesterova: Mm-hmm. Yeah, you can exactly want to. Yeah. Yeah, but do you think with that, like you said, doctors of course they don't want to be replaced? But I also think it's a natural conclusion that if a machine can make a decision or can diagnose something faster or better, with hopefully as least amount of mistakes possible. Do you think the question of doctors and healthcare professionals in general will rise in a few years? Do we need as many of them? Or do you think it's not gonna be a problem? I know it's a theoretical, it's hypothetical, but that's what I wanted to discuss. Or Inbar: I think in medicine... you always need someone to control the decision. It's too risky to let the machine do it automatically. And you need someone to review the decision of the machine learning, the AI. And because there always will be tweaks and specific things for the specific patient that need observation. Anna Nesterova: Mm. Or Inbar: of a physician. is our scene. Anna Nesterova: Mm-hmm. But the way how I see it is also well changing the job of a physician, it will transform it into something else. Would you agree? It's not gonna be exactly the same anymore. They might be more on that side of double checking, making sure AI tools are working well. So what what's your take on that? Do you think their job of a physician will stay the same or will transform it will be transformed? let's go back to CPAT. So specialists right now are specialists because they know how to read the results of CPAT. What's gonna happen with them in the future? Do you think their job will be transformed? They're not gonna be any more tr transcribing and understanding the results. They will be doing something else with AI. So how do you envision that? Or Inbar: It's good question, but if we look at the current situation, there is a lot of mid-level experts in CPED, we must say that, use a few parameters to do the analysis, which is okay, but they can take a lot more information for CPED. So... Anna Nesterova: Mm-hmm. Or Inbar: I think still there will be a lot of people that will know CPED because it's interesting, it's challenging, and they will continue the research and understanding. But this type of platform will help also the mid-level, even to learn more because if you look on the CPED-wise, you can put it in a medical school. And this is actually, you can learn CPEC from the platform. So it will help, think, even to take the mid-level to make them more expert in CPEC. Anna Nesterova: Yes. Yeah. Yeah, that makes sense. And that will also increase the reach and more patients can be seen and more help can be done in the faster process that I can see for sure. I have one more fun question for you. What's one thing about the future of healthcare that you're most excited about? And maybe we can go into something that you're most scared about. Yeah. Or Inbar: Exactly. I will try. What excites me most, I think, is that access. There are incredible diagnostic tools and medical expertise available today, but not everyone can benefit from them. AI has the potential to make expert level care available to far more people, regardless of where they live. And what concerned me most is the possibility of treating AI as a replacement. This is exactly what we're Anna Nesterova: Mm-hmm. Mm-hmm. Or Inbar: talked about before, for human judgment, instead of a tool that supports it. Healthcare decisions affect real life. So we need to make sure AI remains transparent, validated, and always works alongside clinicians, not instead of that this is how I see it. Anna Nesterova: Yeah. And along those lines, while you were training the AI with your product and getting the data, have you seen any signs that AI wants to be independent of what you were feeding it or it's still control you could still control AI and the way how it processes stuff? Or Inbar: Okay, our approach or algorithm, we control it because it's a label, the machine learning and also we can, because the combination of expertise and AI, we can control and see the area that the AI might have mistakes or we have mistake because... Anna Nesterova: Mm. Or Inbar: Even when, for example, my father do a labeling for after five, six, seven tests is getting tired. Absolutely. This is exactly the real life, the real role. But we in Medibite, we can look for that and find the areas that all the experts had a mistake or they, so we control it in 100%. Anna Nesterova: Mm. Yeah. Yeah. I have another controversial question for you. What if let's say AI pick up picks up on some mistakes that doctors were making? Would that mean that doctors were wrong in the past and AI just kind of points out on some flaws? Like it's also very ethical conversation. So what if AI detects that we were doing something wrong? Or Inbar: Yes. you detect different interpretation and I'm sure it's happened a lot specifically in CPET, but in the other end, CPET is not a death or, know, it's not something that can, the interpretation won't create something serious for the patient. So it's a low risk product. It's very, very important. Anna Nesterova: Yeah. Yeah. Mm. Or Inbar: So for sure if you take today's CPET test that is a bit complex to three different physicians that working with CPET, you might not find the same interpretation in three of them. Absolutely. Anna Nesterova: Yeah. I have one more question on that point. So if you are taking information from, let's say, hundred different doctors and you analyze the interpretation, do you train your model to average out the results? Or we have a gold standard that we think it is a gold standard? Because another part of me while you're talking about it, what if one doctor was not doing it right from the start? Or maybe for whatever reason. wasn't correct. But another doctor has an amazing way of doing so. But how do you know and how does AI learn? Or Inbar: When we connected with UCLA, which is a world-known facility for CPED, this was one of the first questions to us, how are we going to validate the product? So... Anna Nesterova: Mm mm. Or Inbar: lucky that there is a process, a regulatory process or clinical process that's been done and for example you take minimal three, first of all the gold standard currently in CPED is experts, there is anything else that you can do but you need to take a real known expert in CPED and you need to take at least three, minimal of three experts to do the study. Anna Nesterova: Mm. Or Inbar: And then there is voting. If three of them decide differently, you need to take it back to the experts. If two decide one, there is a voting. So it's give a weight for the majority of the decisions. And it will be most accurate as can be done in that we hope that In all that we validate, the expert will decide the same, but I believe this is not the case. Anna Nesterova: Yeah. Well thank you so much, Aura. That was a fantastic conversation and I think my audience learned so much about the CPAT and how you training AI. I have one one last question, it's more of a fun one. If you have a wish for AI to do something for you, it doesn't have to be healthcare related, it could be anything. And you have a magic wand and you could A and you could make AI to do something for you, what would that be? Or Inbar: Thank and make me less tired. Anna Nesterova: That's a great answer. I think we all need that. Thank you so much for being on my podcast and good luck with your company, with your project. Thank you. Or Inbar: Yeah. Thank you. Thank you very much.