Latest / Diagnosing Dx / Is AI About to Redefine What a Diagnostics Company Does? Sivan Bercovici, CTO & CBO - Karius
Transcript
- David Filby: Welcome to the show this week, Sivan. How are you? Sivan Bercovici, CTO & CBO: Very good, very good. Happy to be here. Thanks for the invitation. David Filby: Yeah, absolutely. I've been looking forward to it because we've had Alec on recently, but also I like it when we get specific on a topic and arguably there's no hotter topic, but I think we're also at a point when we talk about AI, especially in diagnostics and precision medicine, I think we're starting to see the reality of things. But no one better placed to take us through that today than yourself. So before we do that, why don't you introduce yourself, your background and career and we'll take it from there. Sivan Bercovici, CTO & CBO: Sounds good. Wonderful. So Sivan Bercovici, I'm the Chief Technology Officer and Chief Business Officer at Karius. I've been with the company for over a decade. Here I'm leading the efforts related to the computational efforts and the intelligence powering the Karius platform. We're gonna talk about what Karius is in a second, but just in terms of my background, I'm a computer scientist by training. prior to this deep dive into health and human disease, I was working on microarchitecture and numerical simulation of physical s systems and distributed systems at large scale. But about 20+ years ago I realized that there's a huge role for computer scientists to play in the field of medicine. And that started a journey for me where I initially focused on the study of hereditary diseases, followed by launching a series of companies on rare diseases, then cancer diagnostics, and of course, in this last decade here with Karius as a CTO. My academic background, I had a ton of fun here at Stanford a postdoc, as part of the artificial intelligence group. I retained a research position there to support students in this field, and it was a extremely creative period where it's clear that there's an increasing role that folks can play in moving the field forward. David Filby: How did Karius come about? Because you've been with Karius a long time now. Sivan Bercovici, CTO & CBO: Yeah. So well, similar to many other companies, Karius started actually as research at Stanford. We were studying transplant patients, and this whole field of liquid biopsy was becoming more interesting and more impactful. Interestingly, with the study of transplant patients and whether or not there's a transplant injury, where the idea is that you take a plasma sample and you can track that organ health over time and when we were studying those transplant patients, we realized that there's this compartment of underappreciated biology of pieces of DNA coming from microbes. So it turns out that you and I have microbes on us and in us. And regardless of where they are in the body, when they turn over, similar to human cells, they release the DNA into the bloodstream. And that was the discovery back at Stanford. We realized that there's this compartment that if you can digitize that compartment and if you control for all these different sources of noise, you can actually produce a liquid biopsy for infectious disease diagnostics. You can create a minimally invasive way to study any type of infection coming from any organ without that invasive procedure, without having a strong theory about what's driving the disease. The shift out of Stanford was the idea of converting science into technology, and Stanford is known for creating these companies, and this is exactly the story here. You have to take a observation. Biology that was there always, technology that just became available, the ability to fish out these highly degraded molecules and convert it to something that people will use often that solves an unmet need. So at the intersection of something that was always needed with something that was suddenly possible, Karius was born. We're gonna talk about the AI journey, the intelligence at Karius was there from from day one. But it has been transforming over the past few years and we can unpack that as well. David Filby: That's interesting. And your role is interesting in itself that it's Chief Technical Officer and Chief Business Officer. So how does that work with regards to the intelligence that you guys are driving, and the responsibility is that an indication of just how big intelligence has come in Karius and in the industry that you have those dual roles? Sivan Bercovici, CTO & CBO: Yeah. Yeah, it of course it was very opportunistic and a reflection of the realities that we experience when we engage with life science companies. If you are a cancer company What's often the case is that a pharma company will come to you and say, Hey, we're developing this targeted therapy, we'll need to cover these somatic mutations within these genes. And the diagnostic company, the oncology company, would say, well, here's our platform and we can pick up whichever genes you want and we'll report on whichever somatic mutations you want. There's far more structure in the engagement with pharma when it comes to those sorts of companies. When you look at liquid biopsy for infectious disease diagnostics, the range of requests is far more dynamic. The types of applications people have in mind is extremely wide. And that actually relies heavily on the ability to look at what's needed, thinking about what's possible and creating that bridge. This reflection that you see here, the chief technology officer and the chief business officer, is a mirror of the reality that we're experiencing. In one example, four years ago we got this phone call from our friends at the University of Maryland and they told us, hey, we have this plan to put a pig's heart into human. And we asked when are you planning to do so? They said, Well, it pretty much just happened. We find ourselves in a race to build a unique capability to screen plasma coming from this patient that has a pig organ in them for pig pathogens. On day 20, we found a porcine cytomegalovirus in that patient. And that result has been reported in New England Journal of Medicine, later on in The Lancet and of course, many years afterwards reflecting that result and derivatives of those results. This was January 2022, December 2021, we didn't have a roadmap that says we have to really make sure that we can screen these pigs. But one may say, well, okay, the only thing you're doing is basically identifying more pathogens driving infections. Well there are other cases in which we have pharma companies asking us about chronic disorders. And we've been studying chronic disorders and autoimmune disease and the connection between the microbial compartment, which plays a major role, the human compartment, and all these other conditions. The point is that there's no template or blueprint of how to match the technology we have to what the industry is asking us about. And that's where we're having that dynamic range to go between market need and technology driven possibilities I believe has been serving Karius well. And I anticipate this structure being useful in other companies that are digitizing underappreciated biological signal. Going outside of hey we're studying DNA, hey we're studying ventilation, hey we're focused on cancer. So anytime you're not in one of those multi templated activities, I think that there's a need for that wider dynamic range. David Filby: So you've been at the forefront really of AI in diagnostics for a pretty long time. So talk me through the evolution of AI in diagnostics and how your role has played into that. Sivan Bercovici, CTO & CBO: Yeah. When I started working on problems in human health and disease 20 years ago, the journey was really focused initially on discovery. Can we identify the underlying mechanism of disease in a scale that was not possible before? More than 20 years ago, the studies were extremely biological driven from the standpoint that there was a general understanding of the mechanism and the targeted study of that specific part of biology in the context of particular disease. A significant amount of experiments that produce low throughput data, really focused measurements that benefited tremendously from the intelligence of the scientists driving the study. Over time, we gained access to technologies that allow us to measure biology in far greater resolution, far greater scale, and across an increasing set of alphabets. And by alphabet, measure DNA and RNA and proteomics and epidomics and metabolomics, and you add more and more and more elements of biology and you put a bigger picture together, so big that it's relatively hard down to impossible for a person to actually make sense of all of that data. So the journey really started with discovery when it comes to taking this enormous data sets and making sense out of it, identifying the mechanisms underlying the disease. From there, I progressed to technology development. So, it's great to discover new genes related to end-stage kin disease is one example. But it doesn't really serve the patient. It's a paper. It doesn't influence, it doesn't create the same impact. Converting science into technology means that you need to identify ways to make a product that is relevant, impactful and reliable. And that journey and in my part in that journey requires extremely high discipline. You're not trying to study, 1,000 patients and show general themes and patterns in that data. You're looking at one patient at a time and you're trying to produce a result that will influence their care. So the discipline development of technology is critical and we can talk about intelligence during the age of molecular readout. But that was a second phase. And then maybe a third element has to do with driving utilization. So if the first part was discovery and then the second part is technology, utilization really deals with partnering with those who would benefit from this technology and making sure that they use it in the right time for the right patient. It doesn't serve anyone to have a technology that can do, all sorts of things being implemented in the wrong setting. The outcome is not the outcome people would expect to see and they end up feeling that the process was wasteful. So my role really evolved from as a researcher doing discovery, as a CTO building the technology and now expanding my reach into making sure that this technology is being utilized well. David Filby: Super interesting. So it's evolving with how this industry evolves really and how the AI industry is evolving itself. Where do you see the biggest unmet needs and why does the diagnostics industry need AI? Sivan Bercovici, CTO & CBO: That's a very good question. I touched on that earlier. The first type of AI we've seen was really aimed at digitizing biology and compressing that into a meaningful insight. I would call it the molecular readout intelligence. You're working extremely hard to make sure that you sift through everything you can measure and boil it down to the few things that are actually relevant to an impactful clinical question. And then you build the technology and intelligence around it to make sure that you are reliably reporting on the underlying biology. So for instance, in our case, and that's true for cancer companies and transplant companies, in our case, when we sample and we digitize the signal that is present in the tube, many of these pieces of DNA are indeed coming from the patient, but some of them are coming from the environment. Some of them are coming from the reagents. Sometimes you have collection site contamination events. The intelligence that we embed in the molecular readout phase focuses on the question of what's actually coming from the patient, not from the environment. And what are cases in which the human knowledge is limited and the lack of human knowledge results in a false determination of the cause of disease. And what do I mean by that? Let's say you're the first person who had this new strain of monkeypox. If that is not in a company's database, there's a risk that the machine that is forced to make a call will call a different microbe. E. coli will be called where a person actually has Shigella. Well, that's a problem. So you need to build intelligence into the molecular read out phase to make sure that A, we report only things that are in the patient. And B, we recognize the limitations of human knowledge and we create a system capable of identifying cases in which we're about to make a mistake and stop that system instead of confidently telling you, hey, the patient has X. We're saying, well we actually were not sure. And I would say that was the first era of AI and different iterations of it within the diagnostic field. Compressing biological signal reliably to a focus score or a molecular readout, something that the person can look at, see the level of the signal, say, okay, now because of this, I need to take that action. You we have a track record of innovating as a molecular readout company for over a decade. We are aiming for about 10 different releases a year, about half of those would have impact on the performance of the test, the reliability of the test, the functionality of the test. So we really think carefully about the analytical clinical performance and the utility of the tests. So we have a really strong machinery to produce these releases. Let's see, last month we had one release that focused on bringing antimicrobial resistance results sooner to the physician, not just which microbe we detect, but also what are the resistance mechanisms that are associated with it. And then next month we will release a new version that actually boosts the sensitivity of the tests in key microbes that our customers have spoken to us about. So this innovation engine as a molecular intelligence company is working extremely well. And there's no end in sight. I think if you look at our roadmap and opportunities, I think we're barely scratching the surface of the things that we can do as a molecular intelligence company. But then there's the next phase, the clinical intelligence company, and I'd be happy to talk about that as well. David Filby: Increases in sensitivity. That's super exciting isn't it? Let me ask you this. Karius is a company at a certain phase of its life cycle, you've come a long way, right? And you've achieved some incredible milestones, and I'm sure there's more on the horizon. Are the benefits you're seeing from AI are they hand in hand with the progression of Karius as a company or are they available to early stage companies who are still in their infancy and in development? Are they gonna be able to apply AI like you are or is it or is that something they had can dream of down the line? Sivan Bercovici, CTO & CBO: No, perfect. It's clear that you've been doing this for a while. I think that it's an incredible question. And we are looking at it from the perspective of what can we do next? What are the things we couldn't do before? And it is true, at least in our field, that with the aggregation of additional data, we get to see patterns that were not available before. So if you have 10 samples or 100 samples and super intelligent people in your team, there's only this much you can do. There's only this much you can study those data sets and come up with better techniques. We have reached 150,000 cases. And along the way, between from the first four-sample process in December 2016 to where we are today, we learned a ton. We learned about patterns that may occur in the reagents that we use in order to produce the result. That point in that signal will result in a false call touching a report influencing a patient. We also noticed over time that there are patterns in the data in sick people. Patterns that would suggest that there are classes of microbes that while they are in the patient, they're actually less relevant for the focus infection. These patterns you can only study them when you get to tens of thousands of cases. And these cases cannot be people with infectious diseases. They're different when you're immunocompromised or immunocompetent. It's different where it's a pediatric case or an adult case. It's of course different if it's a case of pneumonia or it's a case of brain infection or endocarditis. The ability to study and appreciate these patterns only happens at scale when you have the right metadata to allow you to do that. So of course we have smart people working here at Karius. There's super smart people elsewhere as well. So we don't have a monopoly on intelligence from that standpoint. But what we do have early access to is a greater amount of data from a type that others find it hard. There aren't labs out there that at this scale can appreciate highly degraded molecules. There's a fair amount of labs out there that can apply strong protocols, highly optimized for the cancer patients or for the study of human hereditary diseases. I think we have a very healthy industry on that front. But Karius specifically is well positioned to study the highly degraded molecules, many of which are coming from the microbes. A fair amount is coming from the human compartment as well. And because we have access to this type of data and at this scale with the type of metadata that we have, our R&D roadmap is super long. And it includes things that it will be very hard for others to imagine are relevant. They only became relevant and realized the minute we actually had a chance to look at the data and say, well, that's strange. That moment doesn't happen without data. David Filby: That's super interesting. So you're at the point now where that large data set is compounding and accelerating its hockey stick because of the data you've got, it's allowing you to learn more, build more, develop more, and it's multiplying all the time. Let me ask you this question. So you mentioned Molecular Intelligence Company. I've heard this a number of times that data ends up becoming the most important thing in this industry. I was going to ask a question, but you answered it in that your R&D roadmap is now being extended because of the opportunities you're seeing. So, as a company, do you see that compounding effect of the data and how you're applying AI to it? Do you see the opportunities to do more in development and go different directions or do you possibly look at your data set as a valuable asset that you start to commercialize access to have a company? What do you do? Do you own it or do you sell it? Sivan Bercovici, CTO & CBO: Yeah, I think this is a super interesting question and I definitely think it merits attention. But the part that I get super excited about is not necessarily how you package it and whether you sell it to others, but I'm so excited about the opportunity that is in there. Myself and my colleagues here, we look at this data and we can clearly see the type of insights and technology that can be developed that will help people that need this the most. We can talk about flywheel effects and getting data and who owns it. And I think all of these are important questions. But I really think that what gets me excited is the realization that you thought maybe that you're done, but no, there's just a lot to build. I think a lot of the transformation that is happening. There is a strong molecular readout world map. We are very likely and very confident that we will continue to develop better and better versions of infectious disease diagnostic tools. The things that we see today makes me extremely confident about our ability to do so. The second shift that is happening though is the clinical intelligence, beyond the molecular intelligence. Molecular intelligence is dedicated to the question of what's in that. Can you tell me what's in that? It doesn't tell you much about what to do with it. What are the implications? And you've seen that trace after context happening in multiple molecular readout companies. We've seen Natera and we've seen Guardant and we've seen CareDx, we've seen Tempus, we've seen Karius. There is this journey that aims to go outside of the normal identity of a molecular readout company. Here's your somatic mutations, here's your MLD results. So the you go outside of that and into clinical intelligence. How can I help you with that patient? And I think Karius, among others, is on that journey. We realize that there's a set of questions we can only answer if we partner up with a physician to get a little bit more context. I will be able to serve you better as a physician if you tell me a little bit more about your patient, because then I can dig into my data and show you other patients in the same situation. When you take a step back and you think about intelligence, it's most appreciated in moments of elevated uncertainty, when the stakes are high, where the impact is meaningful and the burden is overwhelming. That's a basic equation, has nothing to do with healthcare. This is where intelligence is most celebrated. And diagnostic companies, until recently were just adding more information for you, telling you that this information would really help you make the right choice. But as more and more companies are adding more and more information, the need to synthesize that becomes overwhelming. In Karius, we have MDs that have seen 10,000 cases. They just saw these cases. They had the ability to appreciate the connection between the molecular readout, what they learned about the patient from a call from the treating physician, and what the outcome ended up being. How can we make that intelligence available to the many? How can we take things that are a second ago out of reach and make them trivial and abundant? So the journey for clinical intelligence has two components. It has this notion of getting a better appreciation of the context about the patient and then serving intelligence to the physician, not as a displacement act, but as a support, an educational tool. Hey, given everything you told me about the patient, here's information you might find useful. Here's some synthesis of that information. Karius is going through that journey. I'm among the people that are pushing this direction, thinking carefully about our identity as a company, going from molecular readout to clinical intelligence. David Filby: That's super interesting. So we talked about the past and how you evolved with AI and what you did earlier with AI and what you're doing now. What's the difference now then for you in terms of modern AI and diagnostics versus and what's different compared to previously? Why are you so excited? What's different this time round? Sivan Bercovici, CTO & CBO: Yeah, I think that the addiction that our team had a second ago was just to make sure that we give you the most reliable molecular readout. There was high intelligence involved, both from the creators but also the machine. And I wanna make something clear, there is this perspective and I think society has this embedded in their brain that the best intelligence matters. Human behavior and human outcome and the Turing test was all about not being able to separate between a human and a machine. But we're not trying to build planes that fly like human. That would be ridiculous, right? Intelligence is experiencing in a certain context, when you have an objective and when you have choices and when you have uncertainty. But the that intelligence its goal is not to mimic and meet the person's intelligence. It's to aid the person in ways that are beyond reach for people just because of the massive amount of data. So up until a second ago, when AI was called machine learning, the focus was to make sure that you have a reliable readout. It allowed people to deal with the burden of figuring out what to do next. And there are three major moments of uncertainty for those that actually are sitting in front of the patient. The first moment is to figure out what's the right time for what kind of test? And people have their own experience and they have their own biases and whatnot, but that's a huge moment of uncertainty. A second moment of uncertainty, results are coming back. What do I do now? Given everything I know, what's the next step? And then the third moment of uncertainty is to look at the aggregate data as a whole and figure out as an institution, as authors of guidelines, how should we guide the provider, how should we guide institutions to put better policies in place to make best use of certain technologies. So to answer your question on how things shifted, we applied a strong amount of intelligence to make sure that we have reliable call. Now we're shifting towards a better appreciation of the patient as a whole, supporting that journey of the provider when they care for that patient by helping them identify moments in which technology could be most useful. What to do with or without and how to refine any of their their policies. So intelligence is shifting that way. That type of intelligence was out of reach, 5-10 years ago. And now again, it's becoming a more abundant. This is not to say that you don't need to be disciplined in how you implement intelligence. There are principles that Karius believes in when we develop such systems, but that shift is very clear. We are transforming ourselves from a molecular readout company to a clinical intelligence company. David Filby: A big topic that gets floated around obviously is the hype, right? You've now mentioned a few points in terms of the reality of the AI improvements that we're seeing, but where's that the hype versus reality in diagnostics? Sivan Bercovici, CTO & CBO: Yeah. You it's funny, sometimes hype is hype, sometimes hype is being rediscovered as a new reality. There's quite a few evolving beliefs. And some of these beliefs are seasonal in terms of what people believe right now and now it changes a second later and changes back again a few minutes later. And what's super cool to see in some talks is people get on stage and talk about what they were promoting, six months ago and what is their current belief now, and then you actually track them over time throughout their talks and you can see how I wouldn't say they're shifting with the hype, but they're totally shifting with their belief. There is one belief today that says hey you want greater intelligence we need more compute we need more parameters we need more inference time and clearly there are results supporting that. There's a counter belief that says well for a given job it's not that I need for any moment in time, the best position that can also fix cars and can do your taxes. all these, here's your mega brain trying to help you. But I just need, a small model here that will tell me what are the entities that are mentioning a piece of text. In other words, smaller models that are empowered with significant amount of context, but you don't need a nuclear reactor to power them. They can run on your phone. So there is this tension between large models, large power consumption, more compute to be that angle versus small models, more context maybe can run on the edge. You I would say that right now my read is that there's a little bit more hype on we need more compute. Maybe it's true. How to tell. maybe another hype is that there's two extremes. All jobs are being replaced, and then versus we'll all be fine. Following economists, it's interesting to see that some of them have converged into a point that they're not taking sides anymore. They're saying, well, here's one trajectory, here's a second trajectory, here's a third trajectory, and here's data that would tell us that we're this one versus that one versus that one. Before it was a little bit more polarized. And you would hear comments around, well, I'm looking at this self-driving car. How was the last Uber driver augmented by this new AI? where is the augmentation of that driver? And to some extent, here's a person that they is no longer doing that job. Maybe they can get upskilled to do something else. I don't know if you can actually say it, but I think that the two extremes are being hyped. The we'll all be fine or all jobs will be replaced. The honest answer I think is we don't know how things would evolve. Again the last thing I think is being hyped is really putting a person as a benchmark, and how we are making good progress. I think when you benchmark against human you're creating a distortion on how this technology should actually evolve. I mentioned that we're not trying to build planes that fly like human. we're not trying to imitate human flight. Humans cannot fly. They can only fly through these tools that they've built. I don't want to build a calculator that does the math like human. It will be very wasteful. And actually it will be probably slow and it will sometimes make mistakes. So I think that we need to as a community, we need to think carefully about the roadmap for intelligence, the development, the benchmarking, the principles of deployment of intelligence, how to measure progress, when to know that we're getting off track. Benchmarking against human task. There's a ton of emphasis on that. Not that there's something wrong about understanding that trajectory, but we're missing this huge piece that speaks to things that humans just can't do. I cannot see DNA. I cannot just pick up a thing and see DNA. I can look at a file, still can't do anything with it. I think that there's a huge set of opportunities that are not trying to emulate what a human would do in that sense. When I look at a pile of DNA data as a human, don't imitate me. I cannot produce the value. As a scientist that thinks about algorithms, maybe. But there are machineries there that will be extremely well served if we think about problems not from the standpoint of how a human expert would do it, but from the standpoint of how we would design machines that would get to whatever objective we had in mind. So I think there's a ton of focus on imitating human and to a lesser extent focus on solving a problem. David Filby: That's super interesting. And I think on the job side, there's some crazy theories out there, right? And I think Sivan Bercovici, CTO & CBO: Yeah. David Filby: that there's some big statements by really important people which should be listened Sivan Bercovici, CTO & CBO: Yeah. David Filby: to. But I also think there's a job market that's actually showing that the productivity gains are being reinvested in people. And that's a good thing that companies still want great people in their company. Some people have overshot it and now they're bringing people back and they're still using people at the core of it rather than just, hoarding profits and stuff like that. They're seeing the opportunity to grow again. You mentioned mistakes in there and you and I talked offline about standardising mistakes, but where do you see regulation and ethics come into it and the risks associated with this application to the industry? Sivan Bercovici, CTO & CBO: Yeah. These are super important topics. Let's break it into several parts here. In terms of regulation, I think we had plenty of practice when it comes to molecular readouts. How do we validate what this machinery does and running the clinical studies to it to show that there's clinical validity and clinical utility for whatever molecular dealt you have. And we are applying similar mechanisms for these intelligence-based tools. Things that are going one step further and towards the patient, towards the case and towards synthesizing multiple sources of information. And you can see these, different communities popping up here and there trying to think carefully about the principles behind those implementations. Anywhere from transparency, how do these machines work, how were they evaluated, what are the limitations to observability. I wanted to be able to look at how an outcome happened, given some action taken, given some decision a physician made, given some input they got from a machine, given some input that the machine had from its sources, we need to have some some audit trail. We need to think carefully about how the tools doesn't put doctors on autopilot. The intent of intelligence isn't to say, we really don't want you to think, here's the answer. There's another perspective that says, hey, I think that this synthesis would help you appreciate some nuance here that will help you treat your next patient. I think that if you package tools and you design tools in such a way that the doctor just got better even when the machine is off, then you really have a big window. There was a study in which people providers gain and loss access to AI tools in the context of endoscopy. And it showed that, prior to any access to AI tools, they had a certain performance. After access to AI tools, they had almost equivalent performance, maybe slightly lower. But then on days that the machine was off, they couldn't access the machine, actually the performance dropped. So it's not exactly the skilling as losing attention to some extent because the got used for the machine to think for them put them on autopilot so the discipline deployment of intelligence the principal deployment of intelligence is critical you can wing it. I think regulatory wise we're getting better and better at identifying what needs to get done to say here's the performance here's the reliability here's how we validate it. Having said that there will be a moment in which it's more likely that a physician will make a mistake that the machine wouldn't. I think as a society we're very sensitive to a car accident happening to a self-driving car, even if the statistics show that the rates are are low. We just have a different bar. David Filby: Yeah. Sivan Bercovici, CTO & CBO: We feel very comfortable firing a model quickly because of one moment of disappointment. We don't do that with human beings. So we are clearly setting a different bar. As a society we feel uncomfortable getting the same results for a machine. If the machine is not perfect then it's not ready for prime time. I understand the hesitation. The problem with machines is when it has a systematic error, the scale of applying that error is massive. If one person or one physician does one mistake, that mistake doesn't propagate, to a million people due to some nightly release of a software update. So I understand the shape of the risk. Then the other element of if we move from regulatory to ethics, I think that a lot of the breakthroughs that we had were due to perturbation in the data. Meaning that people are noticing that when this happens, you get this outcome. I think that there are some risks when machines are being tuned to apply a certain policy. meaning that they they observe certain things and they act in a certain way. And if they apply it again and again and again, the only thing they can deliver is what they were trained on. Their ability to break off that performance curve is becoming more challenging. There needs to be some perturbation. So when it comes to ethics, there are questions around are we pushing society to a plateau in terms of performance by deploying, one or two or three machines to make medical decisions. On the other hand, there are ethical questions around not allowing access to these machineries because we feel uncomfortable moving some of these decisions to machines or getting physicians to get more support. There's a future state in which maybe it's actually in areas that don't have easy access to better care, that they'll experience actually better care due to the leveraging of AI in areas that a second ago had the best care in the world, because of hesitancy around leveraging such tools, patients have lower access to better care. I think the questions there are clearly questions there. David Filby: That's really interesting. Because I guess the worst side of that would be that the populations of people that already have the best Sivan Bercovici, CTO & CBO: Yeah. David Filby: care and with the accelerating rate of AI as we spoke about earlier, seeing great care become even better whilst Sivan Bercovici, CTO & CBO: Yeah. David Filby: people that don't have access to great care at the moment fall behind further. That would be really sad. Sivan Bercovici, CTO & CBO: That's one trajectory, but it could be that it's actually the opposite. That David Filby: Yeah. Sivan Bercovici, CTO & CBO: in areas where people have a harder time getting access to better care, those areas will be more open to the implementation of, more aggressive, more advanced clinical intelligence. and actually in areas where care is excellent, there'll be hesitancy to move as fast. David Filby: Right. Sivan Bercovici, CTO & CBO: And so it's really hard to tell where things will go, but there is a version in which areas that have the worst access will progress the fastest. David Filby: That's super interesting, isn't it? The stubbornness of existing robust processes and workflows that might be hard to change. Sivan, this has been super fascinating. I feel like we're scratching the surface and we could probably talk for hours, but we don't Sivan Bercovici, CTO & CBO: Yeah. David Filby: have hours. But thank you so much for coming on and shedding a light on on the reality of today and the future of how AI is being deployed into diagnostics and the excitement that I can feel the excitement from the gains and momentum that you guys are seeing, but hearing about how that's being applied into your, R&D outlook, that's super exciting as well. So thank you for coming on and sharing the insights. This was fascinating. Sivan Bercovici, CTO & CBO: Yeah, thank you for inviting me.