Scaling AI in Healthcare: Why Deployment Is the Real Challenge

Stewart Gandolf (Healthcare Success): Welcome to the Healthcare Success Podcast. Today I am pleased to announce or to introduce Dr. Nina Kottler. She's Chief Medical AI Officer for Mosaic Clinical Technologies today. First of all, welcome Nina.

Nina Kottler (Mosaic Clinical Technologies): Thanks, Stewart. Great it's great to be with you today.

Stewart Gandolf (Healthcare Success): I'm glad to talk to you and there's so much to unpack today. So before we dive into the podcast, I wanna set the stage a little bit. You're a practicing radiologist who joined Radiology Partners when it was a startup, right? You were the were you the first, if I remember that correctly?

Nina Kottler (Mosaic Clinical Technologies): I was the first, which is very strange now that we have 4,400+ radiologists. So they tend to call me Rad One. and it was at the time just the two co-founders. So I was the very first employee of the practice many, many years ago.

Stewart Gandolf (Healthcare Success): That's fantastic. And boy you I'm glad you said yes to the job, right? What if you'd

Nina Kottler (Mosaic Clinical Technologies): Yeah.

Stewart Gandolf (Healthcare Success): Say no?

Nina Kottler (Mosaic Clinical Technologies): Funny it's sliding doors right like you never know what's gonna happen and in fact it was during a time in radiology where a lot of people were very depressed about where radiology was. Everything goes in cycles and there were too many rads at the time for the amount of work and reimbursement was going down and although the two co-founders were not radiologists themselves they were the ones that were most excited about taking this opportunity to improve the quality and change the value of what were doing. And I'm like, this is fantastic because I'm all about opportunity. And opportunity is best done when things are hard. And if things are hard now, instead of thinking of it being a challenge, think about it, what you can do with that. And so I was super excited to join them.

Stewart Gandolf (Healthcare Success): That's great. You know, I talk about this a lot. Rahm Emanuel used to say that. I actually use this internally. Never waste a good crisis. Right. If you have that foresight, because it's hard to move people off center otherwise, right? They have to be in pain before they do things differently. So that's great. Fantastic foresight. And I'm glad it worked out for you. Obviously, that's pretty good.

So number one to 4,400 is a big leap. So today you're still working with but as a subsidiary, we mentioned Mosaic a few minutes ago. And you're leading us in AI and leading AI strategy. So I'd love to talk a little bit about, you know, we're gonna do more than just the origin story today, but I'd just like to maybe a little bit about the origin story with Radiology Partners, what it was like then, how things are going on now, then we'll jump into the meat of the podcast today.

Nina Kottler (Mosaic Clinical Technologies): Sure. So RP back in 2013, well, that's when I joined them. They developed as an organization at the very end of 2012. And I met them in the very beginning of 2013. So that's it, 13 years ago now. And back then, if you asked radiology practices what like who's better? Who has a better quality radiology practice? Why would you say that you're better than someone else? They would give metrics that were very operational. Things like turnaround time, maybe peer review, things that didn't necessarily relate to downstream patient care. And there was no sort of standardized metric that everyone used to say “this is what a really good radiology practice is.” And the idea back then is if you can create that, that would be important not only for radiology as a specialty. But also for downstream health care and for patient care. Like let's create the scorecard, the metrics that drive the value that we know we want to create. And if you can't measure it, like you can't do it. So the very first thing were tasked at doing was creating a metric that would define a little bit more about what actual clinical quality was. The second idea with the group was that the radiology practice environment at the time, the average practice size, like how many people in a radiology practice, because maybe you don't know, or many maybe people in the audience don't know, radiologists generally are not employed by the hospital. Radiologists are they have their own group that they work together and they're contracted by a hospital.

And the average practice size of a radiology group back then is about 10 radiologists. And there were hundreds and hundreds of these groups across the country and each one doing something a little bit different. And the idea is, well, what if we could scale those groups? If we could make them bigger, we could make them and number one, have more dollars for investment. Number two, take these metrics that are quality metrics and expand them across a very larger breadth of radiology to improve patient care across more of the U.S. And then three or four, could we actually improve the quality overall and start investing because you'd have more money, invest in technology that would make us better and better.

And that was the premise back then. No one knew if it was actually gonna work, but that resonated with me. I love that idea. Let's use that in order to get better. And we were pretty successful.

Stewart Gandolf (Healthcare Success): That's fantastic. Now did you mention PE earlier? Did you have PE that early? Did you have venture capital? How did you guys spend all?

Nina Kottler (Mosaic Clinical Technologies): That was, it was funded externally. So Rich and Anthony, who were the co-founders of Radiology Partners, had experience with that and they had groups that trusted them enough as people who could create a business that they were able to get groups that were interested in having a very long-term output. Like they weren't interested in skimming across the top. I think it was private equity. I think a lot of people, when they hear the term private equity, they're immediately thinking, they're just trying to skim dollars off the top, take money for themselves and their investors. Whatever happens to the organization happens and they'll dump it and get rid of it.

And that's just I think that's just not a fair way to and not even appropriate. How could you group one full set of investors in a single way. And the way that I think about it is you've got to look at what the dollars are going toward. And for us, we put all those dollars toward quality and development and improvement in the organization. So yeah, it is private equity backed. They had dollars before I joined and we have had multiple subsequent investments and we've used those dollars to invest in technology, to continuously improve not only the quality of what we do from an accuracy standpoint, the breadth of what we do, but also the capacity in which the system can do it in. ‘Cause we just frankly right now don't have enough physicians.

Stewart Gandolf (Healthcare Success): It was really interesting, by the way. You know, we a lot of our listeners are from private equity, so they're not they're definitely familiar with private equity. But my distinction was usually I hear VCs taking on new businesses that you private equity usually gets involved a little bit later. So I think that's really intriguing. It's something that you don't hear it very often, but obviously it worked out well.

Nina Kottler (Mosaic Clinical Technologies): Yes, the groups that we're working with, and I hope I'm not using terminology incorrectly, probably should have Rich answer that question because this is his area of expertise. But the group that we first worked with was New Enterprise Associates, and they are the largest equity group that does both healthcare and technology, and they saw very early on the ability to change an entire system and when you're changing a system to improve it for the better, it's not a short-term investment. It's a very long-term investment.

Stewart Gandolf (Healthcare Success): Well, radiology is super capital intensive, as you know better than I do. So and there's some big salaries there as well. So definitely would require some capital to pull that one off. So well congratulations on that. I think it's really intriguing. I guess before we pivot to AI, which is we're gonna spend most of our time today, I would love to know, given that you really have changed the structure of healthcare, and I'd like to know just a little bit about how things have changed. You know, how you work with hospitals, you know, how do you partner with partners or compete, you know, is it cooperation, the landscape, there's more options now. How are things different today than they were back in the day?

Nina Kottler (Mosaic Clinical Technologies): It's extremely different. I've been in radiology for over 20 years and in my career what I've seen is ebbs and flows. And I mentioned in 2013 when I met Rich and Anthony it was an ebb. That was a time where there were too many radiologists and not enough exams and reimbursements were going down. It was a very like difficult time. And generally it was like every four or five years you'd have too many, too few, and you'd cycle.

That's not happened over the last five or so years. We have only been in a cycle where we have too few rads, even fewer rads, even fewer rads, and it's getting worse and worse. There's always been an increase in the amount of imaging that's ordered because imaging as a diagnostic test is actually quite helpful. About 85% of the time, where a patient goes to a clinician and says, I've got a problem and I want to figure out what's going on. About 80% of the time, they're ordering some kind of imaging study to help figure it out.

And it's because it gives us the opportunity to see inside the patient, and that's extraordinary. So it's a very useful test and it's been getting ordered more and more. Now, beside it getting ordered more and more, the technology itself is advancing. So while a CT scan of the brain, so a CAT scan of the head, when I first started, was probably 35 images. Now it could be 100 to 500 plus images. So we're getting more exams, we're getting more images per exam, and that's been increasing for a while.

Were able to manage that by just getting more and more efficient as radiologists, like running faster on that treadmill but keeping up. As of about 2022, so now that's four years ago, we weren't able to keep up anymore. There's only so fast you can go before you get flipped off the back of that treadmill. And so that's what's been happening since then. And what is that effect that it's having? Well, it's changing how we think about the problem.

In the past, I told you we first came into radiology thinking let's improve the quality, let's define quality and use that definition to improve it. Now, quality is not the biggest problem. If you can't even get to the exam for five days or a month or two months, I'm hearing in some cases, then the quality is secondary to just getting the capacity. So capacity's been the bigger problem and that's the biggest thing that we need to solve right now.

Stewart Gandolf (Healthcare Success): So that's a great lead into AI, I think.

Nina Kottler (Mosaic Clinical Technologies): Yeah.

Stewart Gandolf (Healthcare Success): That's really funny. By the way, I had a scan recently and it's like, wait, that's like just a routine thing. It's like, what's taking forever? Now I have some insight why. I was assuming they found something they're scared to tell me, but apparently maybe that's not it. So one of the things we're gonna talk about today really is scaling AI. And during our pre-call, you said a comment that really struck with me that the next challenge isn't a building AI, it's actually deploying and scaling it. why is that so hard? And that's something that would be, I think, counterintuitive for a lot of people.

Nina Kottler (Mosaic Clinical Technologies): Yeah, and it's not because building an AI model is not hard. Building an AI model is hard, but deploying it safely and effectively is actually even harder. And people forget about that because we're thinking about AI as the tech, and tech is so complex that's the only problem. But in healthcare, especially, just our environments in healthcare are extraordinarily complex. And unless you're living in that environment, it's really hard to imagine. And if you create an AI tool that is meant for healthcare, it has to integrate into that really complex environment.

And the underlying infrastructure for which most of healthcare is run on was created a very long time ago. And it's not as sophisticated. And there's multiple components. So the environment that I live in radiology, so the imaging environment, digital images are shared through what is called the PACS. PACS stands for picture, archive, and communication system. And that system was developed.

The tech for that was developed in the 1980s, 1990s. So it's pretty old. We didn't have the modern AI coming out back then. Then there's electronic medical records, there's archives for data, there are different naming definitions. There are things that are in all different components of the hospital and getting them to connect together is very hard. So first of all, you're putting AI in that environment.

Then beyond the technology itself, once you are able to deploy it, you have to integrate it into the system of record. So you want to make sure it's usable by the clinician. And if you're not integrating it into the workflow of the clinician, it just makes it harder. And actually that's what's happened with most of radiology AI right now. And I will tell you, radiology is further ahead in healthcare than anyone else in AI. Because AI came out in 2016 in radiology.

So we're further ahead, and yet it's still really difficult because the AI isn't fully integrated into our systems because our systems were built 30, 40 years ago. The other piece is that once you are able to integrate it, you still have to now think about AI in a little different way. AI is not just a technology tool, it's a clinical tool. AI is helping clinicians and it's working with clinicians. That means you can't just train people on the buttons and how to use it and what the user interface is going to look like. You have to train people on how it's going to work, get them to understand when it works well, when it doesn't work well.

Because what do you need the clinician to be able to accept whenever the AI is right and reject whenever it's wrong. And that is not obvious. And a lot of people haven't learned about AI, they don't understand it. So there's a whole bunch of components that make it a lot more difficult. And the last one that I'll say, because we talked about this in the very beginning, for environments, is scale. Building an AI model at scale is very different than building an AI model that you could pilot in one organization and then expect that it's going to work everywhere else. It actually doesn't. There's a lot of underlying engineering that has to happen to make it be able to do that.

Stewart Gandolf (Healthcare Success): So the engineering is really important, but in other conversations I've had about this topic, it's the human element, right? Isn't that a key issue of getting adoption?

Nina Kottler (Mosaic Clinical Technologies): Yes. So we have very little autonomous AI. Autonomous AI means the AI does not involve a human. It just does the work on its own. And in healthcare, that's just not something that we are comfortable with right now. And the AI, frankly, isn't good enough, especially if it's any clinical decision.

Humans ultimately have to be responsible for other humans. Tools are not responsible for humans. So we don't have a lot of autonomous AI. Well, what does that mean? That means that the AI is working with the user, the clinician. And instead of thinking about how do I just deploy the AI, you have to say, how do I deploy the AI in a way that the human plus the AI are going to be better together?

And you're right, that's very hard. And part of the reason why it's hard is because we are naturally biased by computer systems. If you think about when GPS first came out many years ago, there were all kinds of accidents that happened. People turning the wrong way down one way streets, they're driving into parked cars, there's people that drove into lakes and you're like, how could that possibly happen? And they say, Well, the GPS told me to do it. And you're like, Well, no, that doesn't make sense.

That's not logical. Why would you follow that? It's because it's not a conscious thing. It's an unconscious thing. So we have to think about how do we deploy these in a way that we can manage that unconscious bias.

Stewart Gandolf (Healthcare Success): So that's such a great metaphor. And AI, I use AI now all day in every aspect of what we do almost. And it's not always right, but it can get fantastic if you hone it. It's like sharpening a sword or something, and it gets amazing. But it's usually actually not right. It's usually not right at first, right?

And it's so you may be halfway there, and that but that doesn't mean you give up on AI, right? And the other part of the human side is in any population of humans, there's somebody there's some people who just like change, they're innovative, they enjoy different things. There's other people that are like, let somebody else get the arrows and there's other people like, I'll be over my dead body, right? And you must have that too, I'm assuming. So you have to figure out how to work around that.

Nina Kottler (Mosaic Clinical Technologies): Really good point. Change management is tough. Like no one likes change. And I always say, like, physicians and radiologists like it even less. And it's not because we're a different kind of human or anything. We're just an environment that is very stressful.

And when you make a change in an environment that's stressful, it causes extensive more stress. The reason why we're able to do some very high level things while we're manipulating our systems is because the manipulation of the systems themselves, we've been doing it for years and it's ingrown. It's kind of like when you ride a bike, right? Like you can jump on and ride a bike and also think at the same time. How can you do two things at once? Well, because riding a bike is rote. You've you've built that into your system.

There's a part of our brain called the basal ganglia. And that part of the functioning is being delivered by your basal ganglia. And that leaves your frontal lobe, which is your executive thinking, open to do stuff. And what happens is, and that's what like when I use my PACS system, that's just coming from my basal ganglia. I have my frontal lobe to think about the patient and what to do. All of a sudden you make a change in your system. It's like changing the gear on like the handlebars so that when you're turning right, you actually turn left. You have to think a lot. You've got to then engage your frontal lobe on that. So how much is left to manage the patient? It's scary. And no one wants to harm anything.

So that's why I say, you know, change is even harder in this role. So you absolutely have to do that. And that requires a lot of change management and education, understanding the AI, providing transparency about how the AI is working. So we're not just making blind decisions, and then monitoring to see how that human AI system is working so that we can go back and show people like, look, here's a mistake, let's learn from that mistake and re-educate.

Stewart Gandolf (Healthcare Success): That's a great metaphor. And I think about that a lot with the unconscious driving on a freeway at 75 miles an hour. And we stop and think about it, like, that's so scary. Like I have people's lives in my hand, including my own and my family's, and it's all on autopilot. And so that's a great description. So on the scale side, you know, we talked about the complexity of doing this at all, but what are the advantages of scale when it comes to deploying something like AI? What are the unique advantages that you have that would be harder to learn at a smaller organization?

Nina Kottler (Mosaic Clinical Technologies): I think well, I gave you some context of what scale is, right? Like scale for us is more scale than most, but you know, going from a practice of 10 to a practice of 4,400, that's a big scale. We do about 10% of all the imaging across the U.S. is interpreted by us and we also do interventional. So that's a lot, that's big scale. But I would say even at smaller levels of scale, scale is important. Number one, because it can help you invest. We have invested probably about a quarter of a billion dollars in AI, which if I were a 10-person practice, even a 100-person practice, like that's impossible.

If I didn't have some kind of dollars coming in to fund this and to allow me to or allow our team to create AI that doesn't give an immediate return on investment, because we're an early adopter. Just couldn't do it. So investment is number one for scale. The second piece for scale is especially this new kind of AI that everyone's using. And I say new, it's not necessarily new to day-to-day work. It's new in healthcare.

The foundation models, these general models that we're using today, whether that's an Anthropic or Open AI, Gemini, any of these models, those are foundation models. Well, how do you create them? They're created with a massive amount of data. That's the difference between the new models and the older ones. And if you have scale, you can train models because you can have a massive amount of data. And one of the things that's really important in healthcare is you want a model that you create to be able to work in other environments.

If I work in a hospital and I walk down the street to another hospital and they have different machines, different protocols for how they do their imaging, I can still read those studies. Even though it's a little bit different, as a human, I can do that. AI has trouble. But if you train it on more and more data, it becomes more generalizable. So you can have one tool that can be useful across multiple different groups. The third thing I'll say with scale is it gives a feedback loop.

So when GPT 3.0 came out, there wasn't a huge amount of talk about it. Like no one talks about it anymore, but what does everyone talk about? Everyone talks about ChatGPT. Why does everyone talk about ChatGPT and not 3.0? The difference between the two was a reinforcement learning with human feedback, RLHF. Basically said, what we want ChatGPT to do is to be able to answer questions from humans.

So let's take GPT 3.0, let's give it a bunch of questions, and then let's have those humans answer it and then train the AI on that question-answer feedback. And you don't need to do a ton of that to get it a lot better quickly. Now imagine in a scaled organization, I roll out a tool to my rads, like 1,000, 2,000, all 4,400 of them. I roll out, I've got a lot of radiologists that are suddenly giving feedback. If I then take that feedback and use that to improve the tool, that's the kind of loop that we need to do to get a tool that is good but not clinically useful to great and clinically useful.

Stewart Gandolf (Healthcare Success): Wow, that's another good insight. So when we talked offline prior to the call, we talked about innovations at academic medical centers. This is historically where things are. And it's really funny, Nina, I've had conversations parallel to that since I spoke to you. Like these things all seem to be bubbling up at the same time in different contexts. And so but the idea that's happening now, that you have some unique viewpoints.

Now we're not putting academic medical centers down, but you have a different viewpoint than they could possibly have. I'd love you to share that. Like what are some of the advantages, you know, a 4,400-doctor independent group could have over an academic medical center? At the face of it, that seems impossible, but in fact it seems it's actually real.

Nina Kottler (Mosaic Clinical Technologies): Yeah, it's very strange. I, you know, academic medical centers was always the place we would go to get cutting-edge research. And why is that? Well, because they would get funds and they would have people dedicated on doing that research. And all the journal articles came out of there. That is where we learned new technology.

It's been interesting with AI that it doesn't mean academic medical centers aren't doing that. They still are. But what's happening is they do it in a lab. And many of them don't have a very big deployment of the AI. So they're learning how to use the AI. Maybe they're crafting it to do certain tools, they're evaluating it, the accuracy of it, but they're not deploying it and they're not deploying it at the scale that we're deploying.

And so what we started seeing over time is that were starting to learn some lessons that the academic medical centers hadn't been able to learn because number one, they didn't have the deployments, and number two, they didn't have the scale. And if you don't have those things, remember we talked about at the beginning what's so important about the what's so hard and what's so important is the deployment. That the tech itself, yes, that's hard to create, but it's half the problem, maybe even less. Most of the problem and the things that you learn are in deployment. And I'll give you an example because we validate every AI model before we deploy it. And in the beginning, we weren't creating our own AI models.

Wewere using vendor models. So we got a bunch of vendor models, we test them out, we're like, okay, they're good enough. We got some accuracy metrics and it works on our data, fantastic, has value, we're gonna roll it out. We then monitor the accuracy of the AI over time once it's deployed. And in that scenario, we found about a consistent 20% drop-off between the accuracy that we measured before deployment when it was sort of like in a research environment and after deployment. And that's because there's so many things that you can't possibly measure in research that you can when it's deployed.

And that's really important because tools like this, especially when they're being created, there's no tool that is perfect when you roll it out. In fact, when you roll it out is when you learn so many of the lessons that you need to learn to make the tool better. Whether that's because there's an edge case, and I'll tell you, there's a lot of edge cases. Like you can't predict healthcare. There's a lot of edge case. How do you manage around those edge cases?

How do you find that the you know data isn't even being routed in the right way or we're sending the wrong information? Or maybe much of the time the AI is not even running. You can't learn that in a research lab. You learn it when you deploy.

Stewart Gandolf (Healthcare Success): So much of life is experiential like that, right? There's things that would never occur to you to come up. So parallels. What can healthcare leaders learn from this shift? I mean that's a big deal. Like this today we have people from all kinds of leadership positions in healthcare. What are some of the things that we can, you know, where else might that apply, these kinds of learnings?

Nina Kottler (Mosaic Clinical Technologies): I think we have to learn that we need to make sure that the end goal is not the tech itself. The end goal is how does that tech fit into the workflow? How does it integrate it into all of those systems that you have that are all disparate in the hospital? And how does the end user and the tool work well together? Because ultimately that's the output. The output is the value of the AI output is not the actual AI output because the AI output itself doesn't affect patient care.

Stewart Gandolf (Healthcare Success): Makes sense. So when you got into this, do you feel like you guys had any idea how big of a journey this is, or did you underestimate that? And then what about others? Do they underestimate? I'm guessing you're going to say yes, but I'd love to hear.

Nina Kottler (Mosaic Clinical Technologies): I think everyone underestimates the power of this. I mean, well, actually maybe I'll say both. There were both ends of the spectrum. There were people that were saying they were overestimating and hyping the technology, which tends to happen if you've ever seen Gartner's hype cycle. That's basically talking about the maturity of new technology. It goes up first, really high, and then it comes down.

And in 2015, which was before we even saw AI in radiology or in healthcare. Geoff Hinton, who's the godfather of AI, said, I can tell already, we already know that AI is going to be better than radiologists. And in five years, we're not going to need radiologists at all. So you should stop training them now. So that was maybe that was 2016 actually. So that was a hype.

That's like, no, we need more radiologists now than we've ever needed. And it's 2026, 10 years beyond when he said we'd need no one. So there were some places where there was too much hype. And the negativity that you get from that hype is you cause a lot of fear and you cause people to be worried about losing their job. And when you do that, that makes people not want to try to use this tech. So we got some of that, some of overhype.

And then you also get underhype where people might see the output of the technology, especially the first time around, and they're like, look, it's actually not even doing that much, not even close to doing my job. I can give you a bit of a level set because people probably assume that the kind of technology they're using today is the AI we're using in healthcare and it's not. The AI that is state of the art in healthcare right now for radiology, for imaging, is not general AI. It's narrow AI. What does that mean? Narrow AI is what Google and all the other groups could do like, I don't know, I feel like 15 years ago.

Where it would take a picture and identify what was in that picture. It might be a picture of a cat and it would say, This is a cat. You're classifying what is in the image. Then at least Google could. You could take a picture of a dog and it would say dog. State of the art for narrow AI classification in radiology right now.

It would take a look at that cat and it would say cat. And then the same model would look at a dog and it would say not cat. Like that's the level at note. It's not doing cats and dogs, it's doing like blood in the brain and holes in the lung and cancer and not cancer. So important things. But state of the art right now for imaging, because it has to go through the FDA and for other reasons, is not general AI.

Stewart Gandolf (Healthcare Success): Wow, really interesting. So how can you know, you're a leader at a big organization, how do leaders impact successful rolling out of AI? Like what are the four or five levers as a leader that you can think of that really matter? Maybe there's only one, I don't know. But there's gotta be some things that you've just find that you've learned from this that other leaders can learn from your experience.

Nina Kottler (Mosaic Clinical Technologies): Yeah. My biggest lesson is that these are collaborative clinical tools. And so you need clinical people to be working on them, validating them, educating the end user about them, monitoring them over time, figuring out product improvements. So one, they're clinical tools, you need clinicians to do this. And what is that process that I just mentioned called? The validation, the monitoring, the education, et cetera.

That's called AI governance. And so that would be my lesson number two is that most hospitals are thinking about governance as a group that does this. They're basically putting, you know, maybe financial people and legal people, compliance people in a group and saying, should we do this, yes or no? That's not governance. Governance is about how do you deploy safely a tool that can add value to the patient. And that requires clinicians.

So that's one and two. I think the other one I'd say that's really important is that AI has a life cycle. AI, the model itself doesn't necessarily change over time, but the data coming into the model changes. We had COVID, all of a sudden there was a new disease we never had before. Maybe you get a new scanner, and that scanner as the images are just a little bit different. And so is it going to work?

You have to continuously review the accuracy of the AI over time. And then final lesson I'll say is again another clinician lesson. You can't assume that the end user knows how to use these tools appropriately. You have to teach them, teach them about the AI, teach them how AI works in general. And then you have to monitor how they're using it, not for a negative way of like, I'm gonna slap your hand for doing something wrong, but from a learning mechanism to re-educate people about things they might not be able to see.

Stewart Gandolf (Healthcare Success): Excellent. We talked about earlier private equity is not evil, right? And I have this conversation here a lot of times from various guests we've had before. And certainly our company does about half of our clients at least have private equity backing. But when you get into scaling like you're doing, there you do need money. There has to be capital.

You can't just do this because you're really sincere and you want to do this. You have to have capital. How do you feel is Capital also equally important for this kind of scale. I mean, obviously to build a t for all the capital equipment, but just to even have the wherewithal to scale the AI. Like, how important is that?

Nina Kottler (Mosaic Clinical Technologies): It's essential and it's essential right now, not because it's always going to be a hugely capital investment product or work as a clinician, because if it were always going to require this much capital, it wouldn't be sustainable in healthcare, but because we're early. And when we're early, you're innovating, and innovation requires capital. So the other reason is that right now in healthcare, most AI is not paid for. There's no reimbursement for AI. Especially if that AI is doing some of the things that we as physicians normally do. Remember how we talked about the problem right now is capacity.

We need to use AI to help have more capacity in the system, which means we need our physicians to be more efficient. Well, you can only make physicians more efficient if you take away some of their work. So we are taking away some of their work. And if you take away some work of someone who's getting paid, there's not going to be extra dollars for that. That requires investment. The other reason is because when I mentioned that AI has to work in a very complex environment, it's not going to be automatic.

You can't just put a really strong, like super capable, really intelligent AI system into that environment and say it's gonna solve all your problems. It actually doesn't. Most of the dollars come from building the infrastructure to connect all the pieces so that the AI that has intelligence can get the data it needs and provide the output in a way that it actually makes an improvement on patient care. So those reasons are essential.

I am talking to members of Congress and hopefully to other government officials like CMS soon to just talk about like how do we make that scalable? How do you manage that over time? And a lot of people are talking about well, just pay for the AI. But I think the right way to do this is not just to pay for the output of the AI, it's to pay for the appropriate use of AI and that's the clinician oversight and governance and things that we need to do to make sure it's done safely.

Stewart Gandolf (Healthcare Success): And you brought up something there before we move on. It's so important. Almost every podcast I get into it comes back to reimbursement, how we're getting paid. So that is such a driver, it's a silent driver that unless you're in the business, people just are not familiar with. So when you just said there, I like, we're not getting directly paid for this, but we still have to do it.

And so it'll be interesting to see as things evolve over time. How are patients benefiting directly? I can guess, but I'd like your insights on this. You know, this is pretty obscure topic. We don't This is a B2B business, you know, podcast, not a B2C, but still everybody at the end of the day is still a patient at some point. So how are patients benefiting as well?

Nina Kottler (Mosaic Clinical Technologies): And that's ultimately what we have to look for because, like you said, if there's no payment, there's not a lot of dollars in the system, the AI has to provide a return on investment for it to be like for it to exist because there's no reimbursement for it. And ultimately the goal is to have improved patient outcomes, which is improved, hopefully improved quality. You have to improve the efficiency of the workflow because it's not just the patient outcome, it has to do with the patient experience as well. And you have to decrease the cost. And those things as a trifecta are just very difficult to do without technology. So what are we seeing today?

Today, like I said, we're not yet at AI that is doing everything that you're using in your day-to-day life, but we are seeing that AI can identify findings that we as radiologists or other end users might miss. No human is perfect. And it's interesting, you might think that the AI and the radiologists could just identify the same finding. And if that's the case, well then how much help is the AI giving you? It's not that much help. What you want is when the AI finds a different subset of findings that are real.

And we are finding that's the case. So the AI plus the radiologists together, you get a much more sensitive and specific, much more accurate report. So you find more cancers, you find more blood in the brain, you find more strokes. The other thing that we're seeing already is that AI, because it's detecting findings, can help us prioritize a study. So if we have a long list of exams that are on red, maybe in some places that I've seen, maybe weeks, what if someone comes in and they're bleeding, like it's an outpatient and we didn't expect it, but they've got blood in their brain. They're just confused and they come in for a head CT.

And we're not gonna see them for nine, we're not gonna get to that report for four days or four weeks or nine weeks. Like that's very dangerous. AI can read through all those studies and they could move that study up higher in the work list. So that patient with a critical finding gets taken care of earlier.

Stewart Gandolf (Healthcare Success): I hadn't even thought of the triage app there. That's amazing. That's really terrific.

Nina Kottler (Mosaic Clinical Technologies): Yep. And more and more important now, the further behind we get. I think the other important thing that we're already starting to see is that AI can take a lot of context. It could go through the EMR and look at prior studies and it could summarize all that information so that it can provide it to us at the point of care while we're making decisions. And non-radiologists who are using AI are using it ambiently. So you can just talk to a patient. And the ambient AI is capturing that discussion and it can create a note for you. So instead of heads down with your fingers on the computer and looking at the computer when you're talking to a patient, you could just be speaking to a patient directly, looking at them and having that real interaction and allowing the AI to do the back end work.

Stewart Gandolf (Healthcare Success): Terrific. so last question for today is what's the future? I'm assuming you're gonna say not that all radiologists will be out of work. So what do you see is happening in terms of the field, in terms of the technology, you know, and are you excited about that?

Nina Kottler (Mosaic Clinical Technologies): I'm extremely excited. I'll start there. I am excited because we are at a point where, again, it's a difficult time in healthcare and we need massive improvements. And there is more openness for change now than there probably has been in any other time that I've been in healthcare. So that's fantastic. AI is going to continue to get more and more capable, especially as we deploy it and iterate on it.

And it's going to help us get more and more efficient as well. So I do think there will be a time where we are caught up. And some people don't think that. I do. I think we're gonna catch up with AI because we'll be much more efficient. So, what happens when you catch up?

There's two options. Either you catch up and there's no more imaging to do, so you take more breaks or you've let people go and you don't need as many humans to do the work as you did before, or you figure out what other value you could add. And that's what I'm excited about. Like I said, right now, because we need to be more efficient, AI has to replace some of the things we're doing. And I call that efficiency AI. But where we're going to is like the next version of medicine, medicine 3.0.

And that is more of an opportunity. What we are finding already is that there's more information in the pixels of the image than we are extracting as humans. We can start seeing things with AI that will help us predict future downstream care. We can look right now at a mammogram, so that's an image of the breast, and it could be normal. There's no cancer that we see on it, but we can predict what's that patient's short-term risk of developing breast cancer in the next three years. We're going to start to be able to do that for more and more diseases.

We will also be able to predict other things. We'll be able to look at an image that has pathology, maybe it is a swelling of the aorta, an aortic aneurysm. That's the biggest vessel in your body. If that aneurysm gets too big, then it will rupture and people, most everyone would die from that. So you want to detect these things. What if we see an aneurysm in two different patients, same size?

How do I know which one's going to get bigger and which one's not? How do I triage those patients? We can find that information in the images of the pixel. And the final thing I'll say is we can in the future be able to look at a lesion in the body. And instead of biopsying it, going in, opening up the patient, taking pieces of it, looking at that under a microscope to determine if it's malignant or not, we will be able to do that with imaging. Imaging is going to get more personalized, more predictive, and more preventive.

Stewart Gandolf (Healthcare Success): Fantastic. Nina, I knew this would be fun. Thank you for your time today.

Nina Kottler (Mosaic Clinical Technologies): It was great talking to you and thank you.