Open Evidence as an Artificial Intelligence (AI) Tool in the World of Medicine

[MUSIC] We are at the oncology brothers. >> For the last few years, we continue to see 40 to 50 new drug approvals and indications each year in the world of cancer. The rate at which the data in science is doubling is getting shorter and shorter. In 1950s, data was estimated to double every 50 years. In 1980s, this was seven years. In 2010, this was estimated to be 3.5 years. And now in 2020s, take a guess. It is estimated to be doubling every 73 days. Every 73 days, the amount of medicine doubles. Hello, and thanks for tuning in. I'm Rahul Gossane, here with my brother and Gohost, Rohit Gossane, and we are the oncology brothers. >> Rahul on our platform, we've talked about treatment algorithm, recent FDA approvals, conference highlights from oncological standpoint. But today, in my opinion, what we are discussing is one of the very important topics of 2026, role of AI, artificial intelligence. And how can we use this technology in providing the best care to all our patients close to home? Yes, there's quite a bit of buzz around AI will replace physicians. But I'm a strong believer that it's not going to be the case, but it's going to replace physicians who do not embrace AI. And this is not a futuristic conversation. This is happening rather today. AI is becoming part of our everyday life where it's changing our practices, workflow, daily management, getting into info about trials and comparing all that, and much more. All right, we will dive into all that where AI stands today in medical field. We are excited to welcome Dr. Travis Zach of physician by training. And now, a chief medical officer of open evidence. Travis, thanks so much for joining us. Thanks so much for having me. Travis, welcome. You know, I have a lot to learn here with Roads background as a computer engineer. And you're running the show at open evidence. I will let you to take the lead and walk me and our listeners through all this. So let's get started. Travis, there's a lot of LLMs out there, large language models. What is an LLM and what makes open evidence different than other existing LLMs like Chad G.P.T., perplexity or Groc? So I know we have a short period of time, so I don't think I'll do the entire semester long course of natural and irresponsibility in the language models. But a way of thinking about a language model is a way for a computer to understand, ingest, and use natural language, which is what you and I use to communicate. That is not a natural thing for a computer to do, which more like, you know, like to use numbers and binary and things. So language models have been around for a fair amount of time and it's a very broad term around how does a computer understand human language. Large language models is just what's happened more recently where we've spent hundreds and millions of dollars, maybe a billion dollars on that task and training a model to do that. In the end, that is a broad and encompassing thing. But the way we talk about it right now for large language models, it's a system where you can interact with it using human language and the output is often within human language. And that's kind of the broad definition of language models. As you can imagine, because so much of our world involves human language, the ability to, for a computer to do that has really revolutionized how we interact with computers and the ways we can use them. Right, Travis, you summed it up well where we were using code C# Java to automate all these things. Now we're using rather large language models to get where the answers that we are looking for. And from these large language models where they were before, they were spitting out information rather fake information with made up evidence and fake references. But this has come a long way just within about a year time. We have seen this improvement drastically where we are seeing elements not hallucinating as much where we are seeing more accurate information. Travis, with the source of NCCN guidelines for open evidence, along with JAMA, any GM, for our listeners, how do you envision to best utilize this tool of open evidence? Yeah, and so I'll take a step back to just talk about why we've gotten so much better in the systems we've created around language models. Because language models are trained on the entirety of the internet. And what that means is when you ask it a question, it might pull it from the NCCN. It might pull it from a Facebook blog, right? And then your answer looks like a hallucination because it can't find a double-devil to it. So one of the things, the structures we've built around language models is making sure when it answers a question, it can only use specific targeted pieces of information to answer that question. And that's a term called rag. It's kind of a process that was actually a couple years old now. But the most important thing with rag is, and it's retrieval augmented generation is the acronym, is that you've created a walled garden where the language model has restrictions on the data it is allowed to use when it answers a question. Now that walled garden then becomes a very important thing to do right. And if information is required that's not in the walled garden, then the language model won't be able to answer it. But if you put information in there that's inaccurate or wrong or misinformation, then the language model might use that, right? So curation of that walled garden becomes a paramount importance. So at Open Evidence, we do two things that's different than a lot of other language models. One, we have a strict focus on medical information. We don't do anything else. That's all we do. And to do that, we work directly with publishers to license content to make sure we have full text and multimedia for many of those. So we license the New England Journal, we license JAMA, we license guidelines, we license the NCCN. All that license goes within our walled garden to make sure when a language model is answering an oncology question, it is answering it based on those guidelines. The second thing, so I missed two things. The second thing is, well, how do you make sure the model knows what the right reference is, right? There are 40 million references. How do you make sure it's the right one? And that's where all the actual exciting machine learning comes in. So I'm an oncologist, right? Not just a physician, I'm an oncologist. And so I may be biased, but like you mentioned, all those things about doubling time of information, that's true in oncology by orders of magnitude compared to other fields of medicine, are field disciplincerine and changing faster than any other group. And so that ML challenge of making sure you have the most recent up-to-date guideline or most recent up-to-date information is harder in oncology than any other. So our goal is constantly building ML algorithms to make sure we have the right references, so that when the model is answering a question, it's using the right references to do that answer. So clearly what you're saying is, this source matters. Where we're pulling all this via our NCCN and EGM or Gem on ecology, where it's giving that references from matters. So this is good and bad, because a lot of the studies, not everything in oncology as an example is black and white. So we perhaps might miss out on some of the nuances and the conversations that happen saying, well, this data is positive, but what next? But here at Open Evidence, we're seeing that it's also playing with that. It's not just giving you data, it's giving you the context of that. And just an example, in my clinical perspective, you know, there used to be time when I had multiple tabs on my web browser, PubMed, NCCN, up-to-date, and you could spend a lot of time to get that right answer for that particular question. Now, you could just type what you're looking for and you get that detailed answer with that credible reference. And just recently, on social media, we were discussing what's the role of brain MRI affront from metastatic breast cancer? There's no clear consensus. But then take a step further. What do we do for those asymptomatic small brain lesions? And if it was Dr. Sarah Simmons from Dana Farber, who brought up saying, hey, go back to Ask a 2020 guidelines for this. And I was quickly able to type this up on Open Evidence and it gave me exactly what I was looking for. That for small asymptomatic lesions, we could perhaps rely on systemic treatment and plan on repeating brain imaging in six to eight weeks rather than jumping on radiation right away. Travis, I understand how this works and I'm heavily using this today already in my practice. But how do you envision this changing over the next year or two? Is Open Evidence going to be part of my EMR to extract patient information and the data to make those treatment recommendations optimizing outcomes? Where do you foresee this going? And also trying in some of the clinical trials prospects. Yeah, absolutely. So there's a couple of avenues we're exploring. I'll mention that clinical trials kind of as a second piece because I also think that's really important. The first piece is how do we make sure that this is integrated within the clinical workflow both in the EHR but then also just so you don't necessarily have to ask every question and things will be surfaced to you that we can actually identify from the EHR. So we're working with a few organizations on that. It's one step at a time and I'm hesitant to announce because I honestly don't know when this podcast will be released. But we're going to announce a press release very soon with a very large community oncology group to integrate this directly within their EHR with the goal of not just pulling the details about the lines of treatment or allergies but also pairing that with the variance of unknown significance with the somatic alterations so that people know if something news come up or something might be off label but is relevant. And then more importantly to what you're talking about, if there may be trials relevant to that VUS that your patient may be eligible for that obviously are not standard of care yet, but could make them eligible for a trial. So again, the reason I'm on this podcast is I'm a I am CMO, but I'm also an oncologist. So I think about a lot of oncology questions first and foremost. So those are the things I'm really excited about in terms of where we're going. With regard to EHR integration, you know, oncology is one of the places where that's most important, just because of how complex our decision making is, not just in terms of what standard of care, but back to what you were saying of all the nuances of what the quality of life and the patient experiences in addition to what evidence in the guidelines say. When they eat it's exciting times. And I'm looking forward to how open evidence integrates everything, whether that's with EHR and offering those clinical trials role. When you were talking about that MRI question, it took me to the time we're right after my fellowship, where I had these quick numbers that I could go back to. That was the rassing medical oncologist. There was one person I was reaching out to, breast medical oncologist. There was one person reaching out to and now it's rather that open evidence. And going back to computer engineering in my previous life and now seeing all this play out in real life, it's I'm just in all how fast the field continues to change. And we need to embrace this. All right, Travis, to address the biggest elephant in the room. That is often saying is if something is free, then no, you are the one who is paying for it. That is by giving your data, attention, or even behavior instead of the actual dollar amount. Can you briefly touch on the business aspect here of open evidence? How is the data that we are putting in being utilized? And we are seeing drug advertisements on the platform. Is that the main source of driving the revenue and are the answers biased in any possible way? Yeah, perfect. So let's cover all those. So about a year and a half ago, we released this free to took off. You know, word of mouth grow, which is where every day we'd wake up and we were incredible to see how many new doctors were asking questions. So we hit an inflection point where we were at, okay, now how do we sustain this? Right? We can't be on VC money forever. We have to have a business model. And there are two options to go. There's subscription or there's free. And we heard from a lot of community doctors that knew that we were going to pull the rug out from them and start charging a hefty subscription. And we didn't want to do that. It felt like it was against kind of the model of an open platform. So we did go ad supported. Now what that means is you do see farm ads, but it does not mean we are selling data to anyone. And I want to be very clear. We have very clear red lines. We never share raw query data with anyone. And we do not train generative models on the data we're receiving. So what that means is we have partners with societies. We have partners American College of Cardiology and CCN, New England Journal. If they want to understand the questions that are being asked by cardiologists or oncologists, that is actually extremely valuable information for them. We have to do that analysis for that because we've set that red line for us that we are not sharing that data. And so we are at supported. So that means if you're an oncologist, you're going to see an oncology ad. You see them at ASCO as well. I think we're almost kind of we're immune to it at this point. But absolutely. And so our business model is showing the ad. But I do want to be clear, our business model is not taking that and selling that data. Travis, thanks so much for touching on that with all that background. As we start to close, I know there is a lot of excitement of around AI, but also fear around AI as well. Any guidance on how our patients could start using these tools to stay up to date with their disease and their treatment options as well. God, this is the stuff that kills me. We talk about this every few months. You guys are the same. You're an oncology, right? Like even before the Gen AI era, Dr. Google was probably more present in our room than any other, right? Like somebody gets a chance to die. I think this they are going to look anywhere they can to find that information. So you know, now they're going to foundation models where like I said, the retrieval may not be just within the medical literature. It could be on health blogs. And so, you know, we are a physician or clinician focus tool, right? Our goal is that the person that uses us has some sort of medical background to understand the literature. And our opinions always bend that patients do deserve a tool to understand the medical literature, but they deserve a tool that's built for them, right? Like it's not just releasing open evidence. There's so much more that you have to wrap around that in order to provide it to consumers. And then additionally, as a clinician, I'm a little selfish. I would want that tool to help them translate what they learn to me. So it's not just printouts from PubMed or like, here's my chat history. So the challenges we are a small company that's grown very, very fast, but not from a personnel perspective. So every three months, we talk, we would really want to do this. We have to build it from the ground out. We have to support it in a different way. But, you know, I think that patients deserve a place to look up medical information that's trusted just like doctors do. But I do think there are specific features of it that have to be a little different than what we do in open evidence. So the long story is we don't plan on releasing open evidence. As it exists right now to consumers, we have thought a lot about what it would look like to create a visit extender for physicians that help their patients navigate. You know, I would say that this is a moving target. We saw this with the publishing data as well. Be a any big publishing house, any GM, JAMA. And now years and decades later, we're also with our published articles providing this layman summary so that a patient who's looking at it can make sense out of it. I would like to believe that in coming years, we'll see something similar be it through open evidence or something that's strictly dedicated to our patients. I totally agree with you. Someone has to do it. It's just got to be someone who actually cares about it. That's absolutely helpful. Absolutely. And again, at least on our platform, we've always focused on reiterating the crunch standard of care or what's going to change our practice today, rather than focusing on how the field is changing over the next five, 10 years. Travis, before we close here, any final thoughts on why oncologists should pay attention to AI tools like open evidence right now, especially given how fast the field's moving. Yeah, I think a lot of our identity and oncology is wrapped around this idea of knowing every trial that's ever existed and every treatment that's coming down the pipe. And that's a really inhuman ideal to live up to in this day and age. There's a separate part of identity, which I think is more human, which is how to interpret what's currently available and apply it to the unique patient in front of us and do that in a way that takes into account all of our experience in treating many other patients in the past. And my hope with AI is that we get to spend more of time doing that second piece and less of our identity are related to making sure we know every ASCO abstract. When I think about my excitement for AI, it's that we do actually get to move towards a more human medicine than we've been practicing in the last 10 or 15 years. Travis, thank you for your time and for sharing your expertise on this timely topic. It's clear AI will continue to play a bigger role in oncology and medicine, not by replacing us, but by extending what's possible in those 20 to 30 minute clinic was exactly to what you're saying. Hopefully with these tools, we're actually walking them through all this and translating that information that we're learning, be it from OE or all our abstracts and studies for our listeners. Let's do a quick recap to tie this all together from today's discussion. In today's discussion, we Dr. Travis Zach, who's an oncologist and also a chief medical officer of open evidence. We touched on the role of AI in oncology and broader medical practice with this particular tool, open evidence. We started with basics on how AI is changing our practice that is utilization of tools like LLMs, that is large language models to process, synthesize and organize medical information in a way that's usable for clinicians. We also touched on current capabilities from answering clinical questions and summarizing the data to supporting real time decision making process. Wow, I am just amazed how good these tools are getting. The important thing here is that we touched on the reliable source it is using to give that data out to us via NCCN or any jam or jammer oncology. From clinical perspective, it really comes down to integrating AI into our daily practice to help us navigate that complex case or to reassure that our recommendations are backed by strong evidence. This is not about replacing us as physicians, at least not yet. But more so, how we think, decide and communicate with our patients. Thanks for joining us on this episode. Tune back in where we get back to our FDA approvals, conference highlights and treatment algorithms. We are at the oncology brothers.