What if the millions of hours of video captured during medical procedures weren't thrown away? A revolution in AI in endoscopy is underway, turning this once-wasted data into a powerful tool to predict disease and accelerate cures. In this episode, we talk to Matt Schwartz, CEO and co-founder of Virgo Surgical Video Solutions, a company at the forefront of this transformation.
Every year, over 20 million GI endoscopies are performed in the U.S., generating a massive amount of high-definition video that, until now, was simply deleted. We dive deep with Matt on how Virgo developed a seamless system for endoscopy video capture, using a small hardware device and a HIPAA-compliant cloud platform to automatically record, de-identify, and store this valuable data without changing the clinical workflow. This enormous dataset, comprising billions of video frames, is the fuel for Virgo’s groundbreaking endoscopy foundation model, EndoDINO. Matt explains how this model, pre-trained on an unprecedented scale of real-world video, can achieve state-of-the-art results in tasks like identifying anatomical landmarks and scoring ulcerative colitis severity. The most exciting application? Using AI for clinical trial endpoints. Matt reveals how EndoDINO was able to predict with remarkable accuracy (an AUROC of 0.78) whether a patient with ulcerative colitis would respond to a placebo, a breakthrough that could revolutionize how pharmaceutical companies design and run clinical trials, making them faster and more efficient. We also discuss the future of this technology, including the potential to analyze the GI tract to find early signs of cardiovascular and neurodegenerative diseases.
Learn More & Episode Resources:
Introduction
Dr. Tim Showalter: Morning, guys.
Dr. Trevor Royce: Good morning, guys. It's good to see you all. We're going to have a great show today.
Dr. Tim Showalter: We're all tired and post-travel, so we have to keep the energy level up.
Dr. Paul Gerrard: Woo! Colonoscopies! Let's get excited.
Dr. Tim Showalter: Have you had one yet?
Dr. Paul Gerrard: No, I have not.
Dr. Tim Showalter: Am I the only one on this call who's old enough to have had one already?
Dr. Trevor Royce: I'm technically, I think I'm due, but I haven't pulled the trigger yet.
Dr. Tim Showalter: I don't think you're 45 yet, are you?
Dr. Trevor Royce: If you have a family history, you bump it up by five years.
Dr. Paul Gerrard: And the colorectal cancer rates have been increasing for younger adults.
Dr. Tim Showalter: Should we turn this into a public service announcement?
Dr. Paul Gerrard: I think it kind of was already.
Dr. Trevor Royce: I think we're there. Let's get screened.
Dr. Tim Showalter: Let's just hop in. Let's get into it.
Dr. Paul Gerrard: This podcast brought to you by Fiber. Reduces your colorectal cancer risk.
Dr. Trevor Royce: Welcome back to Health Tech Remedy, the podcast where three physician leaders unpack health tech innovations. I'm Trevor Royce, radiation oncologist with a background in AI diagnostics and real-world evidence.
Dr. Paul Gerrard: I'm Paul Gerrard, former physical medicine rehabilitation physician, having a focus on reimbursement policy and market access for AI-enabled diagnostics.
Dr. Tim Showalter: And I'm Tim Showalter, radiation oncologist and amateur squash player, who's a former med device entrepreneur now working on integrating AI into clinical workflows.
Introducing Virgo: Tapping into Underutilized Endoscopy Video Data
Dr. Trevor Royce: Today, we're spotlighting a company rethinking how we capture and use one of healthcare's richest, yet most underutilized data sources, surgical and endoscopy video feeds. We're talking about Virgo Surgical Video Solutions, led by CEO and co-founder Matt Schwartz.
Dr. Paul Gerrard: Every year in the U.S., more than 20 million GI endoscopies are performed. And as part of that endoscopy, a video is typically captured. And this is high-definition video, but that video is not really stored anywhere, just still images are. You have somebody going in, doing something that is on video, could potentially generate a video, could potentially have lots of still images within it, and that's all ignored. So, Tim, what are your thoughts on that?
Dr. Tim Showalter: I think people have realized that that's a lost opportunity. Obviously that video is not going to be uploaded to YouTube for entertainment, but it can certainly be used for innovation purposes to improve health care. That's information that's important. So it could be used for quality improvement, could be used for research, could be used to build objective clinical trial endpoints, to train AI models, to train young physicians, and ultimately to improve patient care. Endoscopy is recognized as highly operator dependent. There's the performing of the procedure and the interpreting of the image. And without robust video data, it's really hard to improve consistency and reproducibility.
Dr. Trevor Royce: Yeah, it's a classic case of real-world data. All this data is collected in the routine course of care. And this is just, when you see it, it seems so obvious, yeah, we should be capturing this. Traditionally all this data would be wasted or tucked away and behind some private walled garden, but here they're collecting it in a very systematic way. So they have the system now producing billions of frames and have established the infrastructure to capture and analyze all this previously untapped data. And without that, the insights just disappear. So the modern era with large file sizes, large data sets, all now with the storage capacity for this and the systems and methods in place to analyze all that data, it's just perfect timing for this.
Dr. Tim Showalter: When you get a report for a patient who's had a colonoscopy, there's some description and, I'm not a GI doc. So when I'm looking at it for a patient, it takes me a while to go through it. And there's just a couple still images and it's kind of dependent on the angle, how clear the lens was. And it does make a lot of sense to me that if you've got a more detailed video, you've got a few frames that you can look at and you can see something as the endoscopist is moving across a finding that you can certainly capture much finer details. So I think that makes a lot of sense.
Dr. Paul Gerrard: Oh, I think it really gets us to what Virgo really does. They have a way of capturing endoscopy video, but up front, it's not just the back end capturing it and storing it electronically. It's, hey, how are we going to automate that and make it feasible for the doctor to do it easily because, as we all know, technology and healthcare sometimes don't mix easily. So part of that is making sure that the equipment, the systems are in place so that there's an easy button for gastroenterologists to start to collect high quality video and capture it so that it can be stored for the backend. If their current tools, whether it's software, whether it's hardware, it doesn't allow, you have to have a solution for that. The easier that solution is, the easier you're going to get the data.
Dr. Tim Showalter: You can't fax a video. So we're still in healthcare in the U.S. in 2025.
Dr. Paul Gerrard: Wait, wait, are you saying we still use fax in healthcare?
Dr. Tim Showalter: Oh, yeah. The dirty technology secret. One obvious point, as we're all aware of HIPAA and the importance of maintaining privacy for patient health information. So, of course, finding the right cloud platform has required Virgo to develop a HIPAA-compliant solution where these images can be stored, processed, annotated. And so that's obviously a required part of this foundation.
Dr. Trevor Royce: Yeah, in a lot of ways, the core or the essence of innovation here is the realization that we don't have to treat digital endoscopies like we're still putting things on video cassette tapes and we have to physically store them. It makes so much sense to just start collecting all these data. And then obviously the secondary question is, well, what are you going to do with all these data? And so they're building this incredible data set that can be used for things like study acceleration, precision medicine, obviously leveraging AI to interpret these. To your point, Tim, instead of having a single still shot in the medical record of an endoscopy, why waste all that information with a complex three-dimensional video that's being captured?
From Data Capture to AI: What is an Endoscopy Foundation Model?
Dr. Tim Showalter: It's a great focus to create that substrate. And of course, the next question is, how can you make it easier to glean insights from that? And so, when we talk to Matt, we'll hear a little bit about EndoDINO, which is Virgo's foundation model for GI endoscopy, which is a really important step for getting information out of that and analyzing it. Paul, you want to break down what a foundation model is for folks who may not be familiar?
Dr. Paul Gerrard: And so, first off, I don't think I've ever been able to get a mathematically precise definition of foundation model, but broadly, conceptually, what it is is an AI model that is pre-trained on a massive amount of data sets. And the idea is that you have this pre-trained model that has figured out somehow to extract the key features of, say, in this case, your videos or maybe clips from videos. And so ideally, when you then go to do specific tasks, you need less data to train the AI model for specific tasks because all that extraction of just key features has already been trained on this large data set. And often building foundation models does not require things like outcome data. And since real-time data matched with outcome data is much harder to come by, it's nice when you have something where you can reduce the amount of that kind of harder-to-come-by data that you need for training and validation.
Dr. Trevor Royce: They've collected over 130,000 de-identified endoscopy videos, and that equates to over 3.5 billion frames. So to your point, you can really build some impressive foundational models with that scaled data. That's really unprecedented scale in GI video.
Dr. Paul Gerrard: Yeah. And then where they've started to apply this is things like polyp segmentation, anatomical landmark classification, and ulcerative colitis severity scoring, even with relatively simple task-specific models.
Dr. Tim Showalter: Obviously, there's some importance for clinical care and standardization. I think where you really see the benefit of that ultimately, as things are applied for moving forward cures at the population level, is applying that to clinical research. So at the upcoming UEG meeting, Matt and the Virgo team will be presenting results from a clinical trial in ulcerative colitis, basically using this technology to evaluate endoscopic outcomes.
Dr. Paul Gerrard: Yeah. And my understanding is that the EndoDINO's AI video features could predict eight-week endoscopic healing with an area under the curve of 0.78 in the placebo arm and 0.75 in the drug arm. And for anyone who's familiar with areas under the curve in healthcare, we're talking close to 0.8. That's pretty good. A lot of the clinical prediction tools we use have area under the curve of 0.6 and 0.5 would be a coin toss. So they're up to something that's quite good.
Dr. Trevor Royce: Yeah, I think what I'm most excited about which reflects myself as a researcher at heart is seeing how this can fundamentally change trial design, reduce sample sizes, accelerate drug development, get these innovations to patients faster through this type of platform in terms of even metrics and how those clinical trials, the endpoints and so forth, how did their systems measure those?
Dr. Tim Showalter: So Matt, who we'll speak with later, is co-founder and CEO, and he previously was at Intuitive Surgical. So obviously part of robotic surgery is that you really see the value of real-time video information for improving surgical procedures. He's translated that into bringing forward what he's doing at Virgo.
Dr. Paul Gerrard: And I think one of the interesting things is that he's talked about this concept of feedback whiplash, this idea that every time a board member or an advisor or investor has a new idea, it can create a new direction for the company or try to get the company to focus very narrowly on one issue. And if you do this too many times, you're pivoting, even if it's not a big pivot, it's always changing direction. And so, he's talked about not necessarily ignoring this feedback, but instead really trying to synthesize it into something that is a cohesive direction for the company. And I think all of us have been around startups long enough to realize this can be a real issue.
Dr. Trevor Royce: Yeah, when I was researching this company, one thing that caught my eye that made me appreciate that they're not just treating this as zeros and ones on the screen, but they're actually appreciating the fundamental medicine behind it and having some of their engineers review things like GI anatomy, endoscopy training, some of the fundamentals of this craft and medicine and translating it to the digital world.
Dr. Tim Showalter: Maybe we could just pause and go around the three of us and just do a quick highlight maybe of the areas that we're most excited about this company and to talk to Matt about.
Dr. Paul Gerrard: At the risk of stating the obvious, capturing huge amounts, I think billions of endoscopy frames creates a massive data set, enables a foundational model for GI and AI enabled endoscopy. The downside to that is you have to manage, curate data at that scale. It's technically daunting. And once you show that you can do it and you're creating a business model for it, you pave the way for others to come in after you.
Dr. Tim Showalter: I'm excited to learn more about the EndoDINO Foundation model. Great opportunity to have all that. If you think about deploying that broadly, it's pretty concerning to ask physicians to go through and try to make sense of all the data and really make use of it for each patient and to draw inferences from it. So I think having that foundation model is a really important step to deployment. It seems really challenging to do in a real world or real time environment in the endoscopy suite for clinical applications. So obviously there's compute limits. There's also just workflow challenges that will need to be sorted out for that.
Dr. Trevor Royce: One thing that I would love to hear about that we haven't really touched on yet is the hardware component. It's interesting that they've developed this little black box that basically sits in the endoscopy suite. And my understanding by doing some reading about it is that instead of having to manually trigger the recording of the endoscopy experience, a lot of that hardware piece will automatically do it as soon as the endoscopy procedure starts. So you don't miss out on video by mistakenly not clicking record, reducing some of the friction to start the recording process. And it's unusual for a large data collector like this to have a hardware component. So I would love to hear from Matt kind of what went into that and what the design behind that hardware piece was like.
Dr. Tim Showalter: I hope everyone sticks around for our full conversation with Matt Schwartz, CEO of Virgo Surgical Video Solutions. You won't want to miss it.
A Conversation with Matt Schwartz, CEO of Virgo
We are thrilled to be joined by Matt Schwartz, CEO and co-founder of Virgo Surgical Video Solutions. Matt has a background in biomedical engineering and product management at Intuitive Surgical, and he's now leading Virgo, a company that's transforming gastrointestinal care through video capture and artificial intelligence. Matt, thanks for joining us.
Matt Schwartz: Yeah, pleasure to be here. Thanks for having me on.
Dr. Paul Gerrard: You started your career at NuVasive and then joined Intuitive Surgical, working on the DaVinci robot surgery platform. Can you take us back to the moment when you realized there was an opportunity and something that probably fit nicely with your skill set to transform endoscopy through video and artificial intelligence? And more broadly, what was the spark that led to the formation of Virgo?
Matt Schwartz: Back in early 2016, I was a product manager at Intuitive Surgical working on the DaVinci robotic surgery platform. And I went out to observe a case at the Mayo Clinic. They were doing a total robotic colectomy using one of our new product technologies. And the procedure ended up being extremely challenging. It took maybe seven or eight hours. And while I was there, I remember distinctly in the moment, I couldn't help but think, here I am at the world's most renowned medical facility. They're using what I think many consider to be the pinnacle of surgical technology. And yet for this patient, even the very best surgical outcome that they could hope for at this point would really leave a lot to be desired.
And I was still pretty early in my career. And I started thinking about what might the next 30 years of my career look like and resolved to find a way that I could work on a technology that at least stood a chance of transforming patient outcomes. I had spent a lot of time in the surgical technology space where things we did were pretty incremental in nature, maybe make things slightly easier for the surgeon, maybe hope for marginally better outcomes.
But I was looking for some sort of path that would potentially allow to dramatically improve surgical outcomes or I think most ideally reduce or eliminate the need for surgeries altogether. And shortly after that, I watched a great talk by Steve Jurvetson talking about deep learning. And this is back in 2016 when a lot of deep learning was focused on computer vision and things around ImageNet. I immediately became hooked and realized that AI was a pathway for this type of technology that I was interested in pursuing. And the white bull went off that there's this treasure trove of endoscopy data that nobody was capturing. And if we could become the source of that data, it would help to eventually power what I saw as a coming AI revolution in endoscopic medicine.
The Virgo Cloud Platform: Automating Endoscopy Video Capture
Dr. Trevor Royce: I love that you got your start on the DaVinci robot world. Tim and I are both prostate oncologists and kind of had a front row seat while this amazing technology and transformation of how we think about surgeries are done. And just as a technologist, it's one of the coolest technologies we've seen come through in prostate cancer. So I love that's where you got your start. And it makes sense that you are now doing this incredibly innovative path of capturing all the massive amounts of data that historically have just been discarded during your routine endoscopy. So I think this is a good pivot into your current product, what your focus is. Would love to hear about Virgo Cloud, what it is, and the vision there.
Matt Schwartz: So with our first product, Virgo Cloud, again, we saw this opportunity to capture endoscopy data that was just being discarded. And historically, endoscopy videos weren't being recorded due to what I think is a combination of technical barriers and challenges and then lack of financial incentives. So the recording hardware that was available was pretty unreliable. It was confusing, difficult to integrate and operate in the moment. And for providers and health systems, there was no pre-established financial incentive to capture and store these videos. On top of all that, you've got compliance and storage requirements, and this just created too many hurdles for anyone to tackle the problem.
As a result, when doctors are doing routine endoscopy, whether it's colonoscopy, upper endoscopy, we even think more broadly about bronchoscopy, even laparoscopic surgery and robotic surgery we're now circling back around to, doctors would maybe save a few low-quality still images into the medical record, but no one was saving the video. You'd turn the system off and the video itself would just disappear. And from a purely diagnostic perspective, when you're making the diagnosis in the moment.
That's probably okay for the vast majority of cases. You see the things you need to see to make decisions and move on. But it does limit some potentially interesting use cases. And we were very motivated by this AI opportunity coming downstream and just felt like we needed to do whatever possible to start this data capture at scale. What's been really interesting is as we rolled out Virgo Cloud and gave physicians this capability to save and access all of their procedure videos across everything they were doing, it did unlock some interesting use cases. And to do that, we just had to lower the technical barriers and the price point, frankly, to make that equation work.
So now Virgo Cloud, we've got a small device. It's similar in form factor to an Apple TV. We'll connect that to any existing endoscopy imaging equipment. So it could be for flexible endoscopy, rigid endoscopy, laparoscopic and robotic surgery. And to make things as easy as possible, we focus on automating the recording process. So once you set up the device, it actually uses some patented machine learning to view the video feed in real time, determine when the procedure starts and stops, and that automatically triggers the recording.
It compresses the video, encrypts it, and sends it off to our HIPAA compliant cloud portal. And that little piece, it might sound minor, just automating the start and stop of video recording, but it's really critical. And again, I saw this as a product manager at Intuitive. I would oftentimes try to get physicians to record videos of new and interesting procedures. And unless I was physically there pressing the start and stop button on the recording, compliance was basically zero. And it makes sense. There's so many things going on in the moment when a procedure is happening, the patient's on the table. No one has the time, not the physician, not a circulating nurse, to be responsible for video recording. And so we really focused early on on that user experience and automating the process, which enabled us to then start recording at scale.
Dr. Trevor Royce: It kind of reminds me a bit when email first came out and we treated it like regular mail and you would move your emails to the trash. And now, obviously, as the cost of data and storage decreased exponentially, you're like, why are we doing this? And it's so much better now that you have all your email. Why wouldn't you capture all these amazing images and endoscopies for future use cases? So I think it's that connection that you guys have made is really brilliant. Tell us a little bit from the physician perspective. Does this change their endoscopy experience or is it fit into their routine workflows?
Matt Schwartz: Yeah, it's designed to not impact any of the normal day-to-day clinical workflow. Once you set up the equipment, it lives in the background. You could honestly be a gastroenterologist at a health system where we've installed and not even necessarily know that the system is there running in the background. And it's, again, designed to be that way, to not interfere with the normal clinical workflow.
But from a clinical decision-making perspective, it has started to unlock some interesting and novel use cases. Candidly, early on, we just worked with physicians who were curious and interested in what they might be able to do with video data. We supported some investigator-initiated research efforts where they wanted to have a video case series to work from. But as we started to record at scale, we're now aware of health systems where they've got Virgo installed in every single one of their endoscopic facilities. So if it's an academic medical center, they'll have Virgo at the primary academic hospital, but also installed at the outpatient endoscopy centers that they own and operate or community hospitals that they own and operate.
And in doing so, they're able to syndicate data from across the entire health system. And so we've actually heard of physicians telling us that they've been able to eliminate the need for repeat endoscopy when a patient's going in for surgery, where historically a colorectal surgeon or a general surgeon might have actually done their own colonoscopy the day of procedure. They're able to just go back and review the video. And there's actually been a use case where a patient was previously thought to be inoperable. They had colorectal cancer, and the colorectal surgeon went back and reviewed the initial colonoscopy that was captured in Virgo, was able to actually visualize the tumor inside the colon, and made the determination that they could operate on this patient.
Those sorts of use cases were not things that we foresaw, candidly. We thought Virgo would be used for things like quality improvement, research, training with fellows. And it's certainly been popular in those applications. But kind of like a home security system, we're finding that you don't necessarily go back and review every single procedure. But when there's something important, it's really nice to have the data available. And like you said, in today's day and age, where you can record basically unlimited video from your phone in 4K, why not record everything? And you don't necessarily want to cause burden by having to go back and review everything. But over time, we've been able to create these tools that make it easier to find exactly what the physician's looking for.
Dr. Tim Showalter: This is also one of those examples of basic technology that's existed everywhere else but healthcare making its way. Another example recently is I was consenting a patient for a course of radiation therapy recently. And at my hospital, we just installed the ability to do electronic consent. And as I was doing it, I just thought how ridiculous this is that we've all been using tap to pay and signing agreements for more than a decade everywhere outside of healthcare. And the ability just to store information, have it available, edit it, update it seems like a basic function for consumer electronics, say. And so, I can imagine like you just sort of empower physicians and health systems with the ability to have data and to store data, and let them figure it out. And that sounds like that's what you're finding.
Matt Schwartz: I'll tell you the number one most common reaction I get when I tell people what I do and what Virgo is all about is that they look at me and they say, wait a second, you're telling me we weren't recording these videos all along? I went in for an upper endoscopy just recently. They didn't save the video. I had a colonoscopy two weeks ago. They didn't save the video. And from a patient perspective, it seems fairly obvious. You're going through a somewhat invasive procedure for colonoscopy. There's intense prep required. Why wouldn't you save that data for a use case where it might become important down the line?
Building EndoDINO: Virgo’s Foundation Model for GI Endoscopy
Dr. Paul Gerrard: I wanted to pivot and talk a little more about the AI projects that you're getting into. Virgo announced EndoDINO, a new foundation model for endoscopy. And you're talking to three physicians here, so we barely know what artificial intelligence is. Maybe it's arguable that we even know what intelligence is, but can you tell us what a foundation model is and how EndoDINO is different from other AI approaches, particularly in healthcare and GI medicine?
Matt Schwartz: You'll definitely see different definitions for a foundation model, depending on where you look and who you ask. But I like what I think is a relatively simple definition where a foundation model is just an AI model that's trained on an extremely large and general data set. It often takes advantage of self-supervised learning techniques, which means you're not necessarily labeling everything about the data. And the purpose of this training methodology is to create a powerful backbone that improves your ability to develop downstream AI applications. So the foundation model itself might not actually do something that you're particularly interested in. It might not make a prediction or classify an image on its own, but it's used as a building block that makes those downstream tasks much easier to accomplish with much less labeled data.
So for endoscopy and what we've done with EndoDINO, this means that instead of painstakingly labeling millions of endoscopy images or entire endoscopy videos, we simply drew on the very large endoscopy video data set that we've captured to build this corpus of unlabeled endoscopy images. We've now captured over 2.5 million full-length, high-definition endoscopy procedure videos, which we believe is by far the largest data set of its kind in the world. When we were training EndoDINO last year, we pulled a subset of that data.
A little over 130,000 videos, the equivalent of over three and a half billion frames of endoscopy. And we've got some specialized techniques to pull out the right mixture of those frames and then train a machine learning model that, again, isn't meant to do any one specific thing. We don't tell the system, this image is of a patient with ulcerative colitis, and this is a patient with Crohn's, and here's a polyp, and here's an ulcer. We just feed in the raw images, and there are these techniques you can use. I think the easiest way to think about it is if you took an image and you randomly cropped out part of the image and you put a little redaction box over part of the image, could you train a machine learning model to fill in the blanks there? And this is somewhat analogous to what's happening in large language models where you've got this task of predicting the next word or predicting the next token. And the same thing applies in imaging. And it turns out you can actually train a machine learning model to do that task with enough data and the right techniques and enough compute power.
And so what you're left with at the end, what EndoDINO is, it is a model that can take a new image and ingest it and output a very powerful representation of that image. It kind of compresses everything it sees into these numbers that you can use for downstream applications. And so we've now shown with EndoDINO that its representations of endoscopy are so powerful that it generates state-of-the-art performance in a wide range of GI endoscopy tasks and is actually even starting to unlock some capabilities that no one really thought were previously possible that can't necessarily do as a human expert with human vision. And that opportunity is really exciting to us. It fits very nicely with our overall kind of product and business strategy. And we're excited to see where EndoDINO is going.
Dr. Trevor Royce: Just a quick clarification, maybe for the listeners. When we think of endoscopy, the first thing that comes to mind, obviously, are colonoscopies. Almost most adults in the U.S. are going to be familiar with that. But there are also many other types of endoscopies that are probably less well-known, at least to the lay public, but incredibly common, upper endoscopies, cystoscopies, and so forth. Can you just reflect a little bit on, are you covering all endoscopies or specifically colonoscopy?
Matt Schwartz: Yeah, it's a great point. We do consider endoscopy as this broad umbrella category for any procedure where the doctor is placing a small video camera inside the patient. We started with a heavy focus on gastroenterology. It turns out it's by far the highest volume specialty when it comes to endoscopy and even dwarfs the surgical procedures from a volume perspective. It can be a little bit tough to get good estimates on how many colonoscopies and upper GI endoscopies are done in the U.S. Each year, but you'll see numbers, something like 25 million total procedures done in the U.S. Each year across colonoscopies and upper endoscopies.
So that has been our heavy focus. In upper endoscopy, just for the listeners, this is where the doctor's placing a scope through your mouth, into your esophagus, down into the stomach, and even down into the first part of your small intestine, the duodenum. And there's actually a wide range of procedure types and indications why that procedure is done. It may be done for screening and surveillance purposes if you've got acid reflux and doctors watching out for early signs of esophageal cancer. And then there's actually these kind of endosurgical techniques where they can use the endoscope to get into the gallbladder and the liver and the pancreas and some really advanced procedures take place.
With Virgo, we've really been focused on just all comers. Let's capture as much data as we possibly can. And one of the ways this plays into the foundation modeling concept is, again, the more data you've got and the more diverse that data is, the better. That's what you've seen in large language models. Folks have been out trying to scour for every last bit of data that's available across the entire internet. We find that within endoscopy, when we train a foundation model using data from colonoscopy, but also upper endoscopy, it gets better performance across both domains by having that diversity of data.
We have started to get pulled into some other clinical areas. We do a fair amount of bronchoscopy video capture today. And so this is endoscopy where the scope is going down into the lungs. We did just also start capturing our first DaVinci robotic surgery procedures, which we're very excited about, actually supporting DaVinci for urological surgery and look forward to getting into other areas like cystoscopy, hysteroscopy. The ENT is another area where we think there's broad application. And our overall principle is if there's video, we'd love to capture it. But we do try to stay fairly focused on one therapeutic area at a time to make sure we're adding clinical value back to the physicians as we record.
Applying AI to Clinical Trial Endpoints in Ulcerative Colitis
Dr. Tim Showalter: That's great. I'd like to maybe shift slightly to some of the evidence that you've been able to generate to demonstrate some of that clinical utility. And I did read that at the upcoming United European Gastroenterology Conference in October, and so we're recording this in early September and the presentation will be in October, but I know that you'll be presenting results of applying EndoDINO to a phase three trial and ulcerative colitis to sort of analyze those images. Can you give us a little bit of an overview of what the study design was there and what you hope to show maybe as a little teaser?
Matt Schwartz: Yeah, absolutely. So maybe I'll start by just sharing as we were training EndoDINO how we evaluated its performance and kind of the standard GI AI tasks. As it turns out, GI has been a very popular area for AI research, I think, because of this visual nature. I think it's still true that there are more randomized control trials of AI in gastroenterology than any other medical specialty, even more so, I believe, than radiology. And there have been these predefined tasks that folks have set up and created benchmark data sets. And some of the key tasks are anatomical landmark recognition and classification. So recognizing what part of the colon you're in, what upper GI anatomy you might be seeing. There's another task of polyp segmentation. And so this is not just can you draw a bounding box around a polyp, but can you actually draw an exact mask over every pixel of a given polyp? And then another task is ulcerative colitis disease severity scoring. There's a scale called the Mayo score that's used to score the severity of ulcerative colitis. And we're able to show that EndoDINO achieves state-of-the-art results in all of these different tasks.
And a couple of things that we found really compelling about that, not just the performance itself, but in some cases, we're able to achieve state-of-the-art performance using much less of the labeled data for fine-tuning than prior efforts required. And also, our foundation model was generalizing to these benchmark data sets that were pretty out of distribution for the data we used for training from Virgo. So we didn't capture any of that benchmark data. It actually came from other countries that Virgo doesn't even operate in. It was captured using totally different systems. But we found that our foundation model had become generalized enough that it could extrapolate to those new data sets.
So that was certainly very exciting, but those are kind of the old guard tasks in gastroenterology. And we always had higher hopes for EndoDINO and what it might be able to unlock. And an area that we've been really focused on is inflammatory bowel disease, like ulcerative colitis and Crohn's disease. So we set up this project to work with some phase three clinical trial data from a drug called ustekinumab. The trial is called the UNIFI trial. Again, it's a phase three trial for ulcerative colitis.
And the thing we wanted to study was could we take a patient's baseline colonoscopy and from that alone predict their outcomes, whether in the placebo arm or in the treatment arm. And in an ulcerative colitis trial, there are a couple of different ways that they measure whether a patient is getting better. One is endoscopic healing, which is basically a panel of physicians will review the colonoscopy video and see if that Mayo score has improved. The primary outcome measure is clinical remission, and that is a combination of endoscopic healing and some other clinical outcome measures. And so that ultimately is the determination, did the patient get better or not? And that's what's used to determine, is the drug efficacious or not?
And what we were super excited to find is that particularly in the placebo arm of this trial, we are able to take a baseline colonoscopia DO, process it with the EndoDINO foundation model, and predict the patient's likelihood of achieving endoscopic healing and clinical remission. For the detailed listeners out there, with endoscopic healing, we're able to achieve an AUROC of 0.78. For clinical remission, it's an AUROC of 0.76%.
And for some context, folks have been trying to do this sort of thing for a long time using other standard covariates, things like what was the patient's initial score or did the patient previously fail biologics? And nothing has really worked. I think the best we've seen in the literature is folks achieving an AUROC of around 0.68 to 0.7, even when using machine learning models, combining all these factors. And so what we now have is this AI tool that can, again, take a baseline colonoscopy of a patient with ulcerative colitis and actually predict their likelihood of getting better just on placebo. And I'd be happy to get into some of the applications of that and why that's so important for clinical trial design. But first, happy to stop there and make sure that all is tracking for use cases.
Dr. Tim Showalter: I can already imagine that this could be potentially really useful as a tool for clinical trials. And I'm just wondering, what do you think are the implications for, say, pharma sponsors of clinical trials or those who are really interested in pulling in data from different sources and collecting centralized endpoint assessments?
Matt Schwartz: Pharma-sponsored IBD trials, definitely the commercial area that we are most focused on with this technology. One of the big challenges in inflammatory bowel disease trials is that the placebo response rate is actually fairly high. And it's not just that it's fairly high, it's that it's somewhat unpredictable. And so you'll have these trials that run and they go back and review the data and find that the placebo response was higher than they expected. And as a result, they're not able to successfully demonstrate that the drug is actually better than placebo. And all of this comes back to your initial clinical trial design and how you power the trial to determine how many patients you might need.
And where we see an immediate opportunity is to leverage this technology in clinical trial design. And there's actually some techniques that are we think very much in line with FDA and EMA guidance around using this type of placebo risk scoring to stratify patients at the start of a trial and do better trial design. The end result is that you can power your trial for the same statistical significance or even increase the powering of the trial or have the option to reduce the number of patients that you might need to enroll in the trial.
Which is a huge deal for inflammatory bowel disease, where a phase three trial might require something like 800 to 1000 patients. And these patients are really difficult to find, the trials are very difficult to enroll, they take three to four years to enroll. And when you talk to pharma companies, it's 10s to hundreds of millions of dollars in spend to run one of these trials. And so if you can run those trials 15, 20, 25% faster, more cost-effective, you can get better results, and by increasing the power of the trial, increase the likelihood that you actually hit your primary outcome measure. The last thing you want as a pharma sponsor is to spend five, six, seven years running one of these trials only for it not to meet its primary outcome, and that's sometimes a billion-dollar-plus drug program down the drain. So we think there are huge implications for clinical trial design with this technology.
The Role of Hardware in a Software-First Strategy
Dr. Paul Gerrard: Really interesting potential application. I wanted to pivot, though, and ask a bigger picture question as we get to a close here. And that is, a lot of your focus has been on software, for example, artificial intelligence and data. But on your website, one of the things that we see that's sort of interesting is you have a box, a hardware device. And we've seen a lot of software companies try to stay away from hardware. But it looks like you guys have embraced it to the point that it's front and center on your website. So can you tell us a little bit about that decision and what that does for your company?
Matt Schwartz: Yeah, you know, we were somewhat, I think, forced into getting into the hardware game early on because the goal was to get data off of these legacy endoscopy systems and into the cloud. And just natively, the systems were not really capable of doing the things we needed in terms of automating video capture, doing it in a secure fashion, directing it outside the network. I like to joke with people that I think we might have picked the hardest thing to possibly build a startup around. We got the company started in 2017 where we're working in healthcare IT. We're dealing with very sensitive data. So we've got HIPAA concerns. We're going to build hardware at the same time. And eventually it's going to be an AI solution. And so it's kind of this confluence of maybe the hardest things you could possibly choose to do. But we just believe firmly that this data was eventually going to be gold and we had to do it. So we built hardware out of necessity.
And one of the things we've been very focused on is keeping the hardware somewhat out of the way from a user experience perspective. So when I mentioned before the small form factor of our device being similar to an Apple TV, it's very intentional. Listeners might not be able to appreciate this necessarily if they've not been in an endoscopy suite or a surgical suite. But real estate in a procedure room is at an extreme premium. There's already all kinds of equipment. There's cables that go every which direction. You've got more people than you'd believe walking in and out of the room at any given time. And so you can't have this new big device that's taking up space. And so we really focused early on on building a solution that could just slot itself right into the procedure suite. We actually built the earliest systems on Raspberry Pis, the single board computers, and have since upgraded. But you'd laugh if you saw the earliest devices that we had. We were 3D printing the cases for the devices. So we go hardware out of necessity, but it has become a real moat for us. Once a device gets into the room, it does get pretty sticky with an health system. And that's definitely played to our advantage long-term.
The Future of Virgo: Empowering Physicians and Predicting Disease
Dr. Trevor Royce: Super, super interesting. It's a great discussion and something for a lot of the innovators and entrepreneurs that listen to the podcast to think about. Would love to wrap things up. Any thoughts on what it looks like for Virgo in the next five years? Any particular innovations you're excited about? Anything else you wanted to cover while we have each other today?
Matt Schwartz: There's really two things that we're most excited about with Virgo. We feel like we made this bet back in 2017 that data was really going to pay off. And I think the AI side of things, especially in computer vision and video-based AI, is now catching up to the data that we've got available. One of the things that we're most excited about is actually putting this foundation model into the hands of physicians. Historically in GI, the AI solutions have been dominated by industry players. There's a popular technology called GI Genius that's built by Medtronic. That's an AI pulp detection system. We actually think our foundation model is going to enable healthcare providers to build and validate their own AI solutions.
And so we're actually kicking off a pilot with a very large GI practice now where they'll be looking to use EndoDINO to build their own AI pilot detection system using their own data that will perform in a very similar capacity to GI Genius from Medtronic. We think there's actually a good chance it will exceed some of the performance characteristics, and it will be, again, trained on their own data in their own patient population and customized to the things they like.
We see this as a pretty different paradigm from industry-controlled AI, where healthcare providers had this interesting benefit where they can build and use their own software without necessarily needing to go through the FDA. If they don't plan on commercializing the software, if they just want to use it in their own clinical practice, the FDA doesn't regulate that practice of medicine. And we think this can enable much faster iteration and growth in AI technologies if you just put it in the hands of physicians and physician practices and give them some tools to build the things that they want and let them use it. So we're super excited to see where that goes and what sort of applications expand from there.
The other thing is we do have plans for training a second version of EndoDINO. We've got now about 5,000 times the amount of data available for training the next version compared to version one. And some ideas around training techniques that, again, we think will unlock potentially superhuman capabilities. One of the things that we're most interested in is when you do an endoscopy, whether it's upper endoscopy, colonoscopy, or otherwise, it's this really unique opportunity for a doctor to see inside a patient. And we think that the GI tract in particular can tell us much more about human health than just looking for colon polyps or early signs of esophageal cancer.
The GI tract is very heavily innervated and vascularized. And we're interested in studying whether you can actually predict other conditions. Can you find early signs of cardiovascular disease, neurodegenerative disease, chronic kidney disease, liver disease, all from within the GI tract. And we think the best way to explore that and to uncover that is by training more and more powerful foundation models. So that'll be very much top of mind as we go into training the second version of EndoDINO.
Dr. Trevor Royce: That is a great point. I'm glad we ended on that. I feel like the gut, the gut microbiome is really at this moment of capturing the nation's attention and being associated with all sorts of diseases. So the timing feels very right for this. Well, thanks for coming on the show today. That'll do it for this episode of Health Tech Remedy. Don't forget to subscribe, rate and share the show. Subscribers are so important to supporting our mission. And we're grateful for you spending your time listening to us in these great conversations with folks like Matt Schwartz from Virgo. Thanks for coming on the show, Matt.
Matt Schwartz: Yep. Thanks for having me. Great chat.
