Latest / Powering the Future Within / Episode 24: "Evidence for Every Patient": How Real-World Data and AI Are Changing Healthcare Access
Transcript
- speaker-0: Welcome to Powering the Future Within, the podcast where we explore the breakthroughs shaping the future of MedTech. I'm Omari Bauknight, CEO of ResidentLake Medical. Each episode we speak with a leader who is driving the next wave of medical innovation, helping to advance technologies that are smarter, smaller, less invasive, and more connected than ever before. This is where the future of health takes shape. Let's dive in. Today, I'm grateful to welcome Brigham Hyde, co-founder and CEO of Atropose Health. It's a company working to turn real-world healthcare data into evidence that clinicians and researchers can use to answer some of medicine's most challenging questions. Brigham brings a fascinating combination of scientific training and entrepreneurial experience. He earned his PhD in clinical pharmacology at Tufts. And previously served as president of data and analytics at Eversana. At Atropost, he's tackling a fundamental challenge. How do we help physicians when the published research doesn't adequately address the patients they're treating? We'll explore his journey, the role of AI in closing those evidence gaps, and what it takes to translate an ambitious idea into something that can make a meaningful difference in patient care. Brigham, welcome to Powering the Future Within. speaker-1: Thanks for having me. speaker-0: It's it's really nice to have you on. This question of how do you address patients' symptoms, challenges that maybe aren't well documented or well researched is such an important topic. And so you began your career as a researcher, right? Earning a PhD in pharmacology and experimental therapeutics. And then you moved on to Cohen as an equity analyst covering genomics and life sciences. I'm really curious, what what pulled you out of the lab and onto Wall Street? speaker-1: Yeah, straight from the lab to to Wall Street is an interesting transition. I loved science, loved bench research. my degree also got me into clinical research. most people with the pharmacology degree go on to, you know, run clinical trials or work at the FDA, things of that nature. I was also in Boston sort of during the completion of the Human Genome Project and the early days of that. and the excitement there was around all that. But the thing that was obvious to me was how important data was going to be. and different modes of it and the ability to find, you know, these answers to things we've been trying to understand from that data. and I actually started a company in grad school. so the company's called Relay Technology Management. It's co founded by me and David Greenwald, who has gone on to do many things, but That's currently at Deerfield Capital. And the simple thought we had was, well, if we can get all this genomic data, could we use advanced modeling techniques, math, if you will, to figure out, you know, what the next drug target should be for a given disease? And so we built originally a company to do that. Now, we won a business plan competition. We had a couple of customers. I was wrapping up my PhD, and you know, David was still in his, and I thought. Well, I need to learn something about business. So where do we go to learn that? We go to Wall Street. I don't know that it's quite a business training, a lot of financial training, and got to learn a lot about how capital markets worked, which has served me really well throughout my entrepreneurial career. But that was my original thought. Let's go learn business. and I worked for Cowan as a cell site analyst, covered life science tools and genomics. So I got to know those companies like Illumina and Life Technologies, Thermo Fisher. also CROs close to home for me and running clinical trials and and even a bit of med tech. I've covered striker and Medtronic and Nivasive and all these different things. So yeah, it w I got there. I spent a couple of years there. It was interesting work. I didn't feel like I was building something though, which is what I like to do. And at that same time, that company we started, Dave had now finished his PhD. And we got some funding to go start relay and and we were off. So I I left Wall Street and jumped into my startup and have been building those ever since. speaker-0: Well, so back back to building and the identification of data being powerful powerful tool within healthcare and on the backs of the genomics project, I can imagine a lot of excitement in terms of trying to explore how to leverage that background in terms of data analytics. I'm curious, you know, I've been reading a lot about healthcare data recently, and the amount of healthcare data out there is just enormous. It probably supersedes any other industry in terms of the amount of data that that's available. I'm curious, as you've kind of built a number of things now over time, what did you learn about what creates value in terms of healthcare data versus maybe what's just noise? I mean, how do we determine what's actually valuable and can actually be useful? speaker-1: Yeah, there's been a couple of phases to this throughout my career. if you think about healthcare data in the the old world would we'd be talking about EHR data, we'd be talking about claims data, you know, looking at what was being paid for in the healthcare system. In the early days of that, even when we were starting relay, you know, really a fight to get your hands on any kind of data at scale. It there just wasn't that much of it. The the high tech act that was passed that sort of introduced electronic medical records as a requirement. that really launched the explosion of this market that we've been in. you know, but in the early days it was difficult to get, difficult to possess, you know, security and and things were more challenging. You had to be really careful about access. you know, then I think so there was like the initial phase of finding the data. That was the hard part. So if you if you knew how to find it, you could generate value. the second stage of it was really as the cloud computing revolution took off. It made data easier to work with, and there was more compute and tooling available. you weren't doing it on a local machine, you could leverage, bigger things. And that meant the types of computation you could do really exploded. And so at that point it was the ability to do analytics on that data, early days of AI modeling, I would say, and produce useful output out of that. was really the skill. And, at that time I was in decision resource group. I have chief data officer there. And then we built our original platform as really a wrapper to Snowflake that we could use to deliver value on top of large data sets. And there was an element that happened around that same time where identity handling became a lot easier. de-identifying data makes it much easier, more scalable to use. number two, being able to link data sets, because there's no one data set, you have to link them together. That became easier to do with companies like Health Verity and Datavant and others that really enabled that the industry to more seamlessly work with data. following that period, you know, I felt things started to go more towards multimodal data. And I founded a company called Concert AI in the oncology rural data space, sort of a fast follower to Flatiron Health at the time. And, you know, the real important part was the the depth of data. And this was around the time that the 21st Century Cures Act was passed, which allowed for the idea of using real world data in regulatory approvals and in regulatory and clinical development in pharma, as well as medtech. and that opened up a new economic opportunity because if you could use real world data to shorten the time to run a clinical trial or a control arm Or you could even produce insights that you know would help better inform development strategy and and even launch strategy for pharma. There was a big economic motivation to that. And so with concert and others, we were really focused on deep data that could best represent the patient. Things like notes and omics data were really coming to fore. The early days of imaging data really started to become of interest. So that was sort of the multimodal stage of the beginning of it. It's now the multimodal stage is now expanded to everything's healthcare data, right? Your Fitbit, your you know, glucose monitor, your you know, there's all kinds of data coming at us, even, you know, behavioral data on apps and and things like that could be considered healthcare data in some ways. So that that's had a run that continues. I think the the stage we've been in for the last several years, one of the reasons we launched Atropos was in that phase, okay, it was you could find the data, okay, you could link it, you had cloud and compute. But you're still doing a lot of manual programming. It's still done a lot by humans. Even dealing with note data, you had human people curating information out of unstructured data. And what's happened with LLMs is really the introduction of the application layer, the agent layer, and the idea that you could speak and talk or type to an agent in human text. And get it to do incredibly complex things with data without you having to code. And our original technology we spun out at Stanford University was known as ACE, as part of the Green Button Project, was a big innovation, a big leap forward in the speed with which we could produce row of evidence. And within Stanford, it had opened up the opportunity of bringing that to patients and physicians in the office, right? So you could actually run a study fast enough for it to matter for a patient visit. And give somebody an answer where one might not exist. So that's been the the phase we've been in most recently is sort of the identification of interacting with these data sets. And I think where we seem to be headed next is sort of so what kind of scale and intelligence is possible from all that? you know, we've answered questions from doctors at the point of care, from researchers in pharma for the last five years. You know, really about 18 months ago, we started to say, well, Why are we waiting for questions? if we have all this automation, like can't we figure out, have an agent figure out what questions are likely to come up and you know, answer them ahead of time. And we started doing that. We've actually right now just announced we've crossed 500 million studies. And to put that in in perspective, okay, there's about 37 million studies in all of PubMed ever. So, you know, to multi-hundred X the existing literature and the evidence that medicine's been based on is a whole new wave of this. And it'll be how do we integrate that level of evidence and intelligence, you know, back into day-to-day practice. What what it really means by to take you through this whole path, right, is when we're in Boston after the Human Genome Project, we're all started talking about precision medicine. Right. You're gonna get a patient in front of you, you're gonna have their genomics, their history, and you're gonna figure out what to do for them and just them. Maybe, you know, a specific tailored treatment for them. Well, you know, 20 years later That's finally starting to feel possible. And and the the loop that had to happen was obviously gathering, collecting the data, making a technology enable to interact with it. And now actually already having the answer to every N of one patient, right? I I know what the evidence is that treatment A or treatment B will work for you. I know what it is for somebody else. And so now having those answers, it's like, how do we have a healthcare and payment system that can digest that? But it's possible. speaker-0: When you talk about five hundred million studies, it's even hard to conceive of, especially as you mentioned when the the what's in the current literature domain is is a fraction of that. So can you give in a just a couple examples of when you talk about five hundred million studies, what does that mean? What would be an example of a study in the Anthropos universe? speaker-1: Yeah, so think of how you set about on doing that, right? You you have to come up with a a contour or or a space to consider and you start coming up with like key questions like, okay, for diabetic, should this diabetic be on metformin or GLP one? And let's let's design that as a question where we're gonna follow the patient for twelve months following each treatment, two arms, right? And we're gonna look at several outcomes. They might be A one C level, might be weight loss, et cetera. We're gonna measure that, we're gonna run statistics on it, and we're gonna determine, is there a significant difference in that population for one drug or the other is one better? We now have a piece of evidence, a study, right? Now, how do you get to five hundred million? That's a lot of questions, right? Well, one of the ways we get there is taking a question stem like that. We call it a STEM, right? So GOP versus metformin. And then we run it in every single p subpopulation that might exist. So it's not just for everybody here's the answer. It's for a 45-year-old female Hispanic with CKD. What's the evidence for her? It's for a 55-year-old white male with arrhythmia. What's the answer for her? And so we're running it in similar patients in each, only in those patients in each subgroup, producing the statistical evidence of whether or not It's metformin of GLP. And so you can turn a STEM question into thousands of sub questions and subpopulations and factor in increasing patient context and detail. if you if you look at what's happening in the LLMs right now, people can upload their medical record to Chat GPT. Great, you do that. You ask a metformin GLP question. At this exact moment, the only answer it can give you is whatever study happened to have been run. It really doesn't factor in your history into it. and it the reason it can is not because the model can't do that. It's because there is no piece of evidence for the model to point at if there hasn't been a study run. The the stat we use, Mari, is only about 14% of daily medical decisions have any high quality evidence behind them at all. Right. And that's that's nobody's fault. That's just the way it emerged, right? We couldn't run a clinical trial on every end of one patient. Research progresses at a certain f pace. There's obviously populations that tend to be underrepresented, women, elderly children, certain demographic groups that just haven't been studied enough. Right. But what if we could fill that gap? What if we could fill in the 86%? Now you have evidence for everybody on what to do for them specifically. And and I think what you're seeing out of the LMs is the patients are starting to demand this, right? if I'm going to give you my medical record, I I want, I want to know if it works for me. that's what people want. Right. So our role in this is to fill that evidence gap, is what we call it, that eighty six percent, and make that that evidence available for use to people who are answering questions. speaker-0: Yeah, well, I mean, you're certainly making the concept of precision medicine potentially actionable and in a very, very real way, which is powerful. And when you talk about these five hundred million studies, who's asking the questions? Is it the the frontline physicians? Is it researchers? Is it atropos itself? Is it some combination thereof? How does it all kind of come together? speaker-1: Yeah, it's it's an interesting process, one we take pretty seriously, because you also want to ask good questions, right? and high quality ones. And you know, we've had hundreds of thousands of questions asked of our our technology over the years. So we start there, okay. You know, what are those questions? We also have a a clinical medical team led by our chief medical officer, my co-founder, Sar Gumbar, who acts as sort of the editorial oversight on all of this. Okay. So what should we do next? And I'll just give you an example. We're in the process of doing some of this. there's user questions, sure. There's benchmark questions, the way people are measuring LMs today. There's also key topic areas why don't we take all the guidelines that are out there and look at the guidelines and sort of the branching of them? Like, you know, they tell you what to do at different stages of cancer, for instance. Well, what if we turn each of those little nodes or branches into a set of questions and then run them in this huge data set, expanding the evidence from Hey, this is generally what you do too. Here's what you do in a specific way for a specific patient. There's other things like taking clinical trials. Clinical trials are really questions, right? We've there's millions of them that have been run. Well, what if we think of those as questions as an architecture? Then we program them into our system and run them with that same precision scale. So we're doing all of that. you know, and we're we're often led by, you know, where I I call it the rule of three. So when one of three things is true, we should produce a study. Number one, when there's high stakes situations for patients, right? think of oncology, of course, here. other areas where, you know, having evidence for what to do could really matter. Number two, when there's just a lack of evidence on a given topic, you know, that's a great place for us to fill in. I'll give you a simple example. Geriatric care, strangely enough, became what the the doctors who treat those patients became one of our biggest user groups. The reason for that is nobody runs clinical trials in the elderly. And they're all multicomorbid and they all take multiple medications. So where better to produce evidence for what to do than than in a group like that? So we'll we'll look at those gaps in the evidence base and identify places to do it. we also number three, if you're following the rule three, is where there's expensive medications or treatment decisions, like where it's gonna cost somebody a lot out of pocket and we really want to know it works. where it's gonna be a fight with the insurance company. those are areas we do a lot as well because those are the economics of healthcare that really matter. do you really need this expensive drug? Or would a generic work just fine? Or the payer saying no to your your therapy recommendation, hey, let's find a way to fight. Right. So like those are three guiding principles that generally guide what we do. And we're leaning on the backs of, you know, the research community, the medical guideline communities that have set some of those goals, but we're now filling in what's missing. speaker-0: So then let's let's say a physician is interacting with the atropos and is asking a very specific question about a specific patient. And, you know, obviously part of the the goal is to be able to personalize the answer. So does the response kind of come back, here's the therapeutic considerations, here's kind of the evidence that supports it, and and sort of gives them the the why behind the what, if you will. Yeah. speaker-1: Exactly. So e every response is starts with a summary and that summary will summarize all the relevant studies. We we contextualize both literature and, existing stuff with the stuff coming from our library. I always highlight when that's done. And we'll point to the key facts around that. Okay, so give them a direct answer. In comparison to general LMs, we try to be very prescriptive about evaluating the evidence for what they're asking. And we actually just published last week. A benchmark we call precision evbench. And a benchmark does is it evaluates the answers we're producing or any other LLM are producing and gives you a badge. It says, Did this answer directly answer your question with good evidence? We evaluate the evidence itself, and it tells them red, yellow, green, you know, should you trust this answer? Which is something that we think is really critical because like people are getting responses from LOMs, they're kind of mealy-mouthed, some sound more confident than others. let's get to the heart of it. Is there evidence for this decision or not? Right. And tell you one way or the other whether that's true. So we use a badging system, we use evalves and benchmarks, so we publish, and all that's around building trust, because I think that's the the key to this. speaker-0: Yeah, I think trust is absolutely critical. And you speak to a lot of physicians and they would say, no, well, this patient showed up and they had consulted Dr. Google before I saw them, or now, they consulted Dr. GPT. So I'm curious in terms of building trust, is that something that you've had to try to address proactively? Is some consternation, concern about artificially quote unquote artificially generated responses when it deals with clinical evidence? speaker-1: Yeah, absolutely. it's been a theme for us throughout. we think on the time scale of, you know, changing the future of medicine. And it has to start with this. It starts within Stanford when we ran it. We actually published several papers reviewing the results before it even came out. There's a decade of peer-reviewed publications on what we're doing and how we're doing it. Secondly, the technology that runs the studies is not it it's deterministic in nature. What that really means is. There is no LLM involved in that step, which means we control exactly how it's being done. It's all auditable and traceable. We even evaluate and score the data quality as part of that. So, like making sure if you're going to produce new evidence, it's produced at high quality, best of breed methods, transparent, citable. One of the measures I have for that is we produce hundreds of thousands of these, you know, hundreds of them. Users themselves have taken the results and said, this is great. I want to publish this. And they've submitted it for publication, and we have a hundred percent peer review success rate when we do that. So that's something we tout. the next thing is, these evals and the badging, all of that's designed to be really critical. The the answers themselves and the studies behind them are very transparent in how they work. And I think these things are all crucial. And by the time this is up, actually, this may be out. I'm actually so we produced the 500 million, took about one percent of the library, and we worked with an external publisher to peer review it. and offer peer review for the content in there. And so we'll we'll publish those findings. You know, early, early days looks really good. I mean, you know, still want to make sure you have enough patients for every analysis. So, you know, like some fail for that reason. You know, want to make sure it's a relevant medical question as you're building this out at scale. So, you know, we we don't expect to be peer reviewing all 500 million, but we'll peer review a portion and then every user They get a finding from us, they can push a button and get it peer reviewed within two weeks by other clinicians and we'll pay for that. So, you know, the peer review system's been an important one. I think this scale is gonna change maybe the way that happens. But for now, like, you know, if you want a d doctor to check this out, make sure you think it's right, we support that fully. speaker-0: That's fantastic. I'm curious if you're at the point now where there's some consideration for people using some of these studies in support of submissions. Are are you at that point? speaker-1: Submissions, you mean on the regulatory side? speaker-0: Yeah. speaker-1: Yeah, we have been used that way by our pharmaceutical clients. you know, I I think we're still evolving, you know, the hope in the early days of the 21st Century Cures Act was that we could use synthetic control arms or external control arms from real world data and it would save all this money and time. That hasn't totally played through. There's some examples in hematology, multiple myeloma where this has been done. But that doesn't mean that the regulators and the pharma companies aren't using this type of real world evidence all over the place outside of headline trials, right? we've been included in submissions that looked at things like trial design. They look at real evidence of on market drugs as a way to set benchmark comparisons. In our system, you know, we we look at twenty one CFR part eleven compliance is the standard. This is how data is maintained in clinical trials, for instance. We have the same type of setup with our system. So auditability, traceability of data, provenance, all of that is available, all critical to regulatory. So we're we're bullish about that. I think, you know, we've only seen it increase over the last several years. you know, does it replace clinical trials at some point? I'm not sure about that. And I'll tell you why. You know, if you look at a clinical trial and its purpose, there's a really important element of, you know, the gold standard being double blind, placebo controlled, randomized. Trial, right? And being prospective, right? And that has a very specific value, but I would actually argue that value gets a little overused because its main purpose is to inform the label, right? it is how you justify the words on the label. And I don't think that should necessarily change. I do think at the same time, you have to acknowledge that those trials are deliberately designed to be narrow, right? And and exclude many patients. Perhaps the ones who are going to receive the therapy. Right. So I think they have an adjacent place. And methodologically, the gold standard is the gold standard. And particularly for labeling, should should stay that way. I do think synthetic control arms can play a role there, particularly in certain settings where running a control arm might be considered not the best thing to do. You know, if have a cancer patient, putting them on placebo is not a great idea. So that there are places where I think synthetics belong in all that, but I also think, when you get out to the real world, as soon as we've got our label decided, how do you know if this is gonna work in the forty five year old female span equals C K D if there were none of them in your trial? How do you not? So speaker-0: Yeah, no. speaker-1: that's the place rural evidence belongs. speaker-0: Well, and imagine that all you had was synthetic controls. then all of a sudden you lose the basis to be able to continue to generate the data on which you can extrapolate. So I I definitely agree with you and would make the argument there should be there there has to be room for both. But what you're exploring in terms of looking at the eighty six percent is really fascinating. And if you take a step back to twenty twenty when Atropose first launched. I think you started as, someone was helping from a financing perspective and then you you stepped in as CEO in twenty twenty two. So I'm I'm curious, what was it about the approach, the technology, the company that said, Look, this is something I need to really step in and and and lead personally? speaker-1: Yeah, it's it's an interesting question. I was destined to do this company, just if you think about what I've been doing. Right. Like, you know, if you think about why any of us get into this in healthcare, you know, there's usually a personal story behind it. You know, for me, like I never understood why we couldn't figure out a way to have precision medicine. You know, like it's human genome projects done, like, why can't we do this? Well, you go through all the reasons we can't do it and all the all the stages of that. And to finally be to point where, you could you know, give somebody that direct answer and inform their care. that's that's the moment we're at right now. And this company's been a big part of reaching that. and it's just it's been an incredibly fun ride. I I was having conversation with a colleague who works at one of the big LLM companies, I won't say who, the other day. And he asked me a question, he goes, Brigham, where do you think we are in the overton window of real world evidence becoming a standard? So overton window concept, right? we know something will happen. The question is, are we two years away, five years away? where are we? And it's a really interesting question for me. I'm living it on the front line. It is being used every day. There are guidelines site roll with evidence, right, including stuff we produced. Doctors are using this at the point of care. But at what point do we go from the place of the classic expert led peer review guideline world to, direct evidence for somebody to use? Two two comments I made to when he asked me this. One, you know, clearly I think the scale is there, the methodology and quality is there, it's trustable. and we care a lot about that part. I I think the other part is w who wants this? Okay. Now certainly there are certain doctors who, you know, like doing this a lot, right? But some doctors are saying, you know what, I have a hard enough time keeping track of the guidelines. I'm just gonna try and, you know, practice as best I can as I've been trained. So what are the catalysts? The catalyst for this, I think. Our patients, number one. Back to uploading your record to Chat GPT. you are going to demand it. And if you get a study that says, Hey, this works for me, you're gonna print it out and walk into the office and say, You tell me why this is wrong. Right. And and that starts back to the rule of three at like, you know, more serious patient situations. You know, we have a a big investor group here, a lot of awesome investors, strategics. At this point, I think everybody's, you know, asks us to run a study or two for a family member, right? That's an interesting signal for me. Yeah. Right. So I think those patient that patient push towards personalization and precision is gonna gonna drive it. I think another one is like we need a precision payment system at this point. And let me explain what I mean by that. We have these things this country called formularies, prior off is a dirty word everybody hates. Even payers don't like having to deal with this. They spend like ten percent of their budget on administering this complex plan that is by its very nature, not designed to give the right answer for this patient and this patient. So if I could produce evidence that tells you if GLP1 or metformin is better for you, why do we need to argue about whether it should be covered? Right? Like we all sides should sit around, look at the evidence. If the evidence is compelling and cost effective, perhaps. Okay, great. Like let's approve this. And so my hope is that we get to the point where we have evidence based prioruth. Okay. We agree. If we run a study and the study says it doesn't work for you, I'm not paying for it. You can pay for it. If it does work for you, I'll pay for it. And the benefit of something like this, this is really evidence becoming transactional in nature. It's the unit of measure, if you will. it's like maybe a valuation on a house or something, however you want to think about it. But like when we get to that point, I think the win will be not only better outcomes in our in our medical care, but it will reduce cost because. If you deny a person who would benefit from a therapy, guess what? They're not done. They're gonna go to a visit. They're gonna slip and fall. They're gonna take an opiate. They're gonna go to the ER. they're going to cost the system money anyway. Not to mention all the effort, cost, and time both sides spend right now arguing about it. The call centers, the documentation, the forms the doctors have to fill out, the phone calls with the insurance company. all that could go away. If we just had a transactional system based on evidence. to determine, whether things got covered or not. So that that's that's my big dream. That's that's through the over to window, perhaps. But for this to really become ubiquitous in a standard, I I think you have to link the two. speaker-0: How do you think about closing the loop? So you you finish the study, comes up with a recommendation. Sometimes people follow it, sometimes people don't. Is there openness in terms of closing the loop? Should recommendations be followed in terms of what occurred? How how do you capture that piece? speaker-1: Yeah, we do that with some of our provider partners. we'll track decisions and what happened and did the evidence work? We published a couple of things on this just in terms of, you know, ROI that comes from this and the way they measure themselves, quality scores, you know, cost of care, things like that. And, you know, it turns out following the evidence is a good idea, at the end of the day. I think the hard part is for the system to do that at scale, right? Like, think about what I'm really saying. I'm saying every doctor would treat every patient differently based on individualized evidence. The system is not built for that necessarily. And so thinking about how you would actually get people to follow through with those decisions and, you know, make that a standard practice in a way that would work for those entities, that's what we work on a lot right now. speaker-0: How would it inform guidelines, right? speaker-1: I think it blows past guidelines. We're very interested in keeping the guidelines like the guideline committees can use this data. Like go for it. You know what I mean? It just the classic way they did guidelines, it was very hard to even add one new step to a guideline, right? Exactly. I'm talking about adding millions, right, of steps to it. So I I think the the world will demand that expansion and you know, those groups could be beneficiary. speaker-0: Right. I'm I'm curious, you know, from the time that you joined or you time you got involved in 2020, six years now later, six years plus, and you kind of hinted at some the changes we've seen in terms of artificial intelligence and now everything seems to be about agentic AI, and obviously you all are on the forefront of that. How have you seen things change and how do you anticipate changes to keep Atropost kind of at at the front end? without getting over your skis in terms of innovation in this space. speaker-1: Yeah. I remember the was it Thanksgiving when ChatGPT was released to the public three years ago now. It feels like yesterday. I think every technology company, young or old, sort of sat down and said, Well, wait, do I have a business anymore? You know, can this thing just do everything? and you know, there's been phases to that, you know, even over the last three years as they've you know, changed and evolved and different models have gotten better, different things. the first thing we saw, even in our own hands, but has been reinforced in the hands of our customers, particularly in pharma, is like the idea that you could just stick a bunch of data into Chad GBT or pick your model and it's gonna be able to produce the same things that we produce, it's not happening even now, right? Like it's generally right now pretty good at getting from zero to sixty percent right, but getting to ninety-five, ninety-nine, it's not there. And there are reasons we know why that is, and why our technology gives us differentiation there. But even with that, you have to assume it's gonna get better, right? That's the only the only way to operate as a CEO in this space is like pick a vector. I think this is when they'll be at this point. Yeah. and if you look at us producing the Alexandria Library, it's great, if we have this lead and this advantage, but we know these things are gonna get better, like let's not sit in our advantage, let's start producing it ahead of time. And that you know, data we're creating, that content, that gives us a training advantage. Right. And one that, keeps us ahead. can I tell you if we're at AGI tomorrow and who knows? Could you just type rebuild atropo into Chat TBG and it'll do it? I you know, who knows? But at the same time, like there's training content, there's technology strategy, and then there's vertical nuance. And I think the other thing we're seeing is like, you know, working with pharmaceutical companies, for instance, there's a lot of nuance to that. Okay. And compliance, it turns out. And other things. So forever and a day, in every technology cycle I've seen, there's always been this take cloud. It's like, okay, well, everything's going to the cloud and AWS are, Microsoft are going run everything. It's like, well, they end up needing vertical players who can apply, vertical specific expertise, regulatory and compliant functions. So this feels different, but maybe not that different if you think about those cycles, right? So our bet is to be a great vertical player that understands a customer base that's hard to understand and can work in a regulatory and compliant way. And, right now that's working really well. speaker-0: Yeah, no, vertical expertise is absolutely critical. And that that alone, I think, sometimes drives alignment with an industry that's difficult to replicate. The other piece you've talked about, which I think is really interesting, is being this kind of independent evidence layer, which, you know, you could imagine this capability being aligned with certain other partners and there'd be a lot of questions. So I mean, for example, if this capability lived within a group that was responsible for pre-auth, for example, that would Drive a lot of questions. So I think your focus on being an independent evidence layer is an important piece. Have you found that to be the case? speaker-1: It is. I think about it a lot because between the three of us, we've been working on this type of topic and problem our whole careers, both from within academia and outside. And we really care that what we're producing is of high quality and trustable, right? And if you think about what I'm saying is just as a source of knowledge, it's gonna dwarf the existing knowledge bases of, you know, literature and guidelines and other things just on size. So I think it's really important that it remains independent. We work with many parts of healthcare in a B2B sense, providers and pharma certainly. We work with technology companies, digital health companies, diagnostic companies, med tech companies. So we always stand the idea of, look, we're just gonna produce the evidence. I'm not gonna produce it the way you want. It's like it's gonna say what it's gonna say. You know, at some point I was joking with one of my colleagues from Stanford who was instrumental early on, Tip Kim. And I I was joking to him. He goes to me, he goes, Bring up me it's kinda like you're building a Ferrari here. you're Enzo Ferrari. you're tinkering to make it like this perfect Swiss watch of a car. And I was like, first of all, I take that as a compliment. And second of all, like I wonder if there should be a foundation that governs this editorial process of this machine, or if we should make some version of our library available more publicly, you know, for the good of humanity. You know, who knows? Maybe Stanford and I put together a foundation and Put some of it out there, of course, with our business still in mind. But, i we we do feel like we take the stakes of creating this very, very seriously and really want it to have the impact. if we could achieve the goal of filling the evidence gap, humanity should benefit. That's the way we'd be like. speaker-0: Absolutely. I I love the the analogy to Enzo Ferrari, who was I think also really known to as he produced the product with such focus and detail, there was almost still a little bit of an air of uncertainty around the technology because he kept something kind of close to the vest. When you talk about working with physicians and clinicians, understandably there's sometimes concerns that they're working with a black box. So how do you how do you open that black box to give them enough to understand? where the the recommendations are coming from when they're looking to work bedside with patients. speaker-1: Yeah, no, I that goes back to the publishing of all the methods, the peer review stuff. But even a an individual study you get from us, it's everything you would get in a peer-reviewed publication, all the raw data files. We even go further. we do something called data fitness scoring because running things on real data is great, but real data is a varying quality, right? and so we literally put like at the top a credit score. You know, we say, like, what's the data fitness score for this question? Zero to a hundred scale. So I think, you know. Transparency is the key to that trust, right? And even that that score, we published how we do it, right? Right now our benchmark that we use is on Hugging Face. They can go test it. So, you know, it's sort of the key to the whole thing is being transparent, creating trust. speaker-0: Yeah. And you've done this a few times, so you obviously understand that as an area you really need to lean into. I'm curious, based on your previous experiences working with real world data, have there been any kind of common failure modes or things that typically stall companies in this category? Again, when we're talking about creating businesses that leverage data, there's a lot of watch outs. And I'm just kind of curious on some takeaways or insights or lessons learned having done this a few times. speaker-1: Yeah, I I think it's business model which has helped back this space. And it's challenging because you know, it in DRG and concert, even ever saw nowhere I was, you know, have to go buy data and then you have to analyze it and turn it into something profitable for your customers, something valuable for them that's profitable for you. And Every single one of the companies that's gone deep on that space, with very few exceptions, really struggles to produce profit because the data costs keep rising. And, you know, if you're gonna be the one creating value, you have to buy more and more data, right? So it's just always been a difficult sort of business model to run. It's something we thought a lot about here. And while we focus more on a consumption-based model, kind of think about like loading your token account in ChatGPT and then run certain questions and I use different amounts of tokens. So, know, I do think that's something we think about a lot because if you're gonna have a big goal and a big aspiration, big impact, like you have to take care of the business and it has to run in a way that's interesting to capital markets in a way that can scale, right? and I I think that's held the the businesses back in this space in a lot of cases. You know, I think Tempest has done a really good job. You know, if look at what Eric and Ryan have done, I mean Combining a testing business with a data now AI business, it's put them in a really exciting position. but business combinations have worked that way at times. You know, IMS to IQ via the CRO into the real world data world. you know, that can work in certain cases. And I think all these companies, including us, have the ability to apply AI to the data that we have. And these great tools created by the hyperscalers that, you know, can create more value. But think you've got to have a business model that aligns with that value you're creating. I think that's the the key piece. speaker-0: Right. absolutely. And that kind of goes to a little bit more on building the moat in space like this, because as you alluded to before, with the development and kind of the pace of development of some of the LLMs and AI systems. You can imagine waking up and facing the situation where you're disrupted. But it sounds like you've been really thoughtful about how you might be able to create votes here. Is it the data access, the methodology, workflow integration, evidence library? Is there something else that you've been focusing on in terms of speaker-1: There those are four big ones on our board. we've been super focused in addition to producing Alexandria on workflow integration on the provider side. So, you know, you're not gonna beat ChatGPT at a consumer app, right, or a direct-to-position app. You're just not. So then if you're working with the businesses like the health systems, you talk to them about where do they wanna see this information presented, right? How does it create value for them? And so We've bet really hard on agentic and ambient workflows. We have a very big partnership with Microsoft on that front. Other partners we use for distributions. I'll just give you an example that's sort of interesting. So we started off with an app that you could ask your question, type it in, you'd get an answer from us, right? Including a new study. Nowadays, on the provider side, you'll have an ambient recording recording the visit. While that's happening, it pings our agent in the background. The agent figures out from the conversation and the patient history what the key questions are, and then produces studies and summarizes studies to produce an answer. And when the doctor swivels their chair over to Epic, the answer and the evidence are already there. So think about that going from a pull service where you're stopping asking a question to a push and it's just coming right through the back. And doing that with our provider partners is actually useful because They want to have oversight over where these information flows go and what goes where to whom and how's it being evaluated. And so I think that B to B integration is really critical. And that's the way we bet both with our technology partners like Microsoft and Databricks, but also directly on the ground with our, health system customers, providers, et cetera. speaker-0: Yeah, no, that's fantastic. I could imagine one of the potential hurdles or criticisms being rolling something like this out and saying, Look, it's great, but it takes more time. It adds more time to patient engagement. I only have so much time to begin with. So it sounds like pretty thoughtful in terms of the workflow and actually maybe making these visits more efficient than less. speaker-1: Yeah, exactly. that's what people want. That's where all the technology is supposed to take us. You don't stop and talk to an agent if you have a question, but if you don't need to and it just shows up at the right moment in the workflow, that's that's where it should happen. I speaker-0: I think there's so many different directions you can take this in in terms of the impact in healthcare. And one of the reasons I like working in healthcare is because something we all can relate to. We all have personal stories about family members being impacted by different conditions. And certainly it feels like that's been part of what's made this such a powerful mission for you. Is that something that sometimes you feel like you get deeply connected to? a product or a research direction or customer relationship that makes it challenging to to move away from just because of personal experience. speaker-1: Interesting question, w when you worked in healthcare for 20 years, you sort of wake up one day, you're like, Well, how did I end up doing this? Yeah. Right with all its challenges. you know, I I think you do have to balance that. Like, you know, for us, like being very mission centric to solving the evidence gap, it part of it's intellectual because like the three of us who co founded the company and and many of the team that that we brought on just fully believe like you can chase every problem you find in healthcare back to this, right? Like now. that's a little bit of tunnel vision. when I was when I was a PhD researcher, I worked on mitochondrial biology and in the heme niche. And, we would joke, we'd be like, Well, in a mitochondria lab, every disease in the world is because of the mitochondria. And so you have to check yourself sometimes, like you know, make sure you're not speaker-0: If you have a hammer you see a nail, right? speaker-1: Well, right. and and I think you wanna make sure when you're doing that, if we're focused on closing this evidence gap, it's equally important to produce that evidence at scale as it is to be able to make it easier for people to get. Right. And what's the best way to accomplish that? And that that's a different side of product, right? We go back to the integrations, like why those are so important is like if you really want to get this out there and make that impact, you know, you gotta focus on those things too. So, you know. that's something we talk about a lot internally as a team and make sure that we're not in an echo chamber focused on one problem when there's other ways. speaker-0: Sure. So let's put on your prognostication cap. It seems like that's some a world that you've lived in for quite some time and maybe something you enjoy doing. speaker-1: So we call our our studies prognostograms. speaker-0: Yeah. speaker-1: reference back to I think it's 1971 publication by Dr. Feinstein at Yale that described this type of report for the first time, you know, and actually called it a prognostigram at that time. so we it as a shout out to the the early days of real evidence we we call prognostic. speaker-0: Wow. That's a nice little tip of tip of the hat. so I'm curious from your perspective, let's go ten years from now then. Do you think that the evidence gap will be so narrowed that we will not have to continue to push on looking at new populations of patients who are poorly represented? Or do you think this is a case where we're going to take down different challenges today and it's only going to expose more challenges tomorrow? In other words, are we just gonna unlock more and more patients? inpatient populations that we are not addressing completely. So how far does this how far does this go? speaker-1: It's like think right now we're finding there's increased demand for more precision, more personalization, even more localization. Like, you know, we're doing some work with the Ohio State University, and like something they cared about a lot was we have all this evidence produced on these national data sets. Like we want to know the answer here. Like at Ohio State, does this work? Right. So run it locally on our data. So think of that library of five hundred million being copied, run again at Ohio State. And there's the Ohio State answer. And so that increasing precision, increasing localization will scale out, you know, in the sort of next three years, I would say. you know, from there, I think in the middle years, more multimodal data, I mean device data, other things that can be used as data inputs to creating evidence, that's going to scale dramatically. and I I can sort of see that coming. I think by the time you get out there, I think the last stage of this is like if I could produce all this evidence, including device, behavioral, genomic, biologic data, right? I think it has the potential to have a dramatic impact on drug development. And one part of drug development is like figuring out how to make a drug that hits a target. Another thing is understanding the biology of disease from the real world data and actually saying, based on this, it's likely this drug might work. I I we've done some work with a a group called Every Cure run by David Feigenbaum, where the basic concept is like, We might have enough drugs already. And what we don't have is we don't know this drug that was prescribed for something else would actually work well in this this setting. And we've we've done some of that work ourselves, but I think drug repositioning, also acceleration of label expansion for drugs, all those things are possible. So wonder what that really means is not only are you gonna get the evidence for what you should do, but a personalized treatment to match with your personalized evidence, right? So I think the ultimate goal is like one patient, one treatment, right? And, you're able to produce that in a tight cycle. So I'd be hopeful over the ten year arc that really do shorten the time for drug development, you know, from a patient in the field, you know, directly to something that will that will help. speaker-0: Be fantastic. I'm curious, as we run more and more of these studies and you do it more regionally and then locally, et cetera, do you envision there being periods of time where the answers or responses are in direct conflict? And if so, how do we arbitrate that? speaker-1: The answers do conflict, right? And I would liken it to this. Let's just say you're being treated at a a rural health center that's understaffed and, you know, doesn't have as many sort of experienced specialists, or you're being treated at my wife that works at Columbia University, Columbia Medical Center. Okay. Do you think the outcomes are different for the same treatments at those two places? A hundred percent. Right? So that's actually real. Right. Now, what are you measuring there if you were to show that one worked better at one place than the other? You're measuring equipoise. You're measuring the value of the ancillary and support services somebody might have, which is something, by the way, the health system is scream when they're like the Ohio State example, hey, we treat patients with these medications and these procedures really well. Better than the one down the road. And those are things beyond just the therapeutic modality you might be considering. And it's how medicine is really practiced. So I just think those are real. And there's also, by the way, population differences. Those things are back to the localization, personalization, precision element, those are really important to know. And frankly, like would matter to the ultimate outcome. speaker-0: So they're different but perhaps they're purposefully different inappropriately so. speaker-1: Well, purposely or arbitrarily, but they are. speaker-0: Right. speaker-1: you know, and you can get into a whole bunch of reasons why that might be true that you'd have different outcomes, right? So we're not here to necessarily say, you know, one is better than the other. I'm just saying like this is the facts, right? Like that's our job. speaker-0: Well, it's an important job. And I think as we continue to explore, imagine and envision how we might be able to take generalized medicine and make it into vo visualized medicine, you're certainly serving a very important sector and one that a lot of people will be paying attention to and will certainly be paying close attention to it. speaker-1: Very cool. speaker-0: So this has been a phenomenal conversation, Brigham. Really appreciate you coming on and spending some time with me today. speaker-1: Thanks so much for having me. Great conversation. speaker-0: And thanks to all of you for joining us today on Powering the Future Within. If this conversation inspired you, follow the show and share it with your med tech community. I'm Omari Bachnight. Stay inspired and keep innovating.