AI in Healthcare: Hype or Help? A Doctor's Honest Take

You know, you get this huge queue of all of these images and if the only thing you can see in your workflow is just kind of the, you know, patient's name or whatever, you have no idea which ones you should look at first versus not. And so by having this AI system that runs behind the scenes and goes through all these images identifying things that might be emergencies, it can fix that order and prioritize them and say, "Hey, let's, you know, we'll flag these three. These have some alarming thing in them. Human, you need to go look at this now to just check if there's really an emergency here. " So, should a radiologist be worried that they're not going to have a job in the future?

What are you most excited about in the AI space these days? If you've picked the wrong problem to solve and you're trying to force AI onto it, that can definitely end poorly. Uh, you need to understand like who's going to benefit from this and why are they going to benefit and also why specifically is AI the right tool to solve this particular problem.

AI is it hype or help? Good day and welcome to the show. I'm your host Pete Cooper and our topic is AI hype or help. Today's guest is Rachel Draus. Rachel and her company Glass Box Medicine has an incredible backstory. She has been involved in medical imaging, various aspects of AI across the health care. She's also a double doctor.

That means a doctor of medicine and a PhD doctor. We go really deep into medical imaging and a whole bunch of other things. This is a fascinating deep dive into AI and healthcare. Let's get into it. Welcome Rachel. Welcome to the show. How are you today?

I'm doing well. How are you?

I'm good. I'm good, thanks. I've been looking forward to this. Um we met uh recently and we had a lot of intersections uh and the type of work that we do and I wanted to spend this time with you going in deep about some of the layers of of our work and how it pans out in the AI world. AI is a big space now and there's lots of specialties appearing.

No one's an expert in AI but there's many people who are experts in parts of AI and I think Rachel you're one of those. So, why don't you introduce yourself and what your business is? It's Rachel Dreos from Glassbox Medicine. Correct.

Yes. Thank you. And uh happy to be here today. Uh I'm a physician with a PhD in AI. I'm the founder and principal consultant at Glassbox Medicine where I do AI strategy and implementation projects across all industries. My specialty is custom models on proprietary data sets, especially in computer vision, natural language processing, and domain specific data sets.

In healthcare, I've worked on projects spanning 2D medical images, 3D medical images, medical video, medical audio, ehr data, and OMIX. So, lots of uh healthcare experience and excited to talk today about AI and healthcare AI.

Rachel, that was fantastic. Um, you kind of fire hosed me there with with lots of stuff. I'm gonna start from what I remember at the beginning and work through. So you said you were a physician. Now a physician is a doctor, right? So you actually were a practicing doctor. Let's start there.

I [snorts] did go to medical school. I was at Duke. I did my MD and my PhD there.

And uh after my clinical training there finished after I I got my MD, the normal thing for me to do would have been to go into residency. But I actually went into the health AI space after medical school. So I am a a physician but currently focused on health AI and uh not actively seeing patients anymore.

So if people don't know how what the word doctor means, it has two meanings. It has a medical doctor who is a physician as you describe, but it's also somebody who's got a PhD. They can also be a doctor, but you actually are both. Is that correct?

Yes, I am both.

It's very, very rare. I don't know anybody else who has both a medical doctorate and a PhD doctorate.

Do you know anybody else like you?

So, actually my husband.

Oh, well, there you go. He's going to come on the show in the training program. Um but yeah it is an unusual degree combination. In my medical school class there were I believe 120 students and eight of us were doing the combined MD/PhD program and around year four most of our medical school classmates graduated and then the rest of us you know we were still around. So in total it was an eight-year combined program and around year six I remember thinking this is actually pretty long.

Yes. Wow.

It was exciting to to wrap that up.

Wow. What an undertaking.

Um, and it kind of makes more sense now you frame it that it was actually a structured program that brought the two things together and and you've continued down the sort of the PhD path in your research area. Um, so there are other people out there in the world that are as rare as you. U, that's a comfort to know.

Yes, I'm I'm not the only one.

Yeah. Yeah. So, why did you choose to go the the PhD route versus the the the practicing doctor route? That's a good question. It was a really tough decision because there were certain aspects of patient care that I knew would not be replicable replicable outside of direct patient care.

And there is something really meaningful about you know directly delivering care to sick people. I really liked internal medicine in an impatient setting. I also really liked emergency medicine. I think if I had gone on to do more uh clinical training, it would have been in either one of those areas. I ended up deciding to focus on tech because basically I had to come to grips with the fact that no one can do everything. I think up until that point in my life, I had sort of kitted myself that oh, you know, if I just work hard enough, I can do everything. And I I reached a point where I realized actually I can't do everything. And I had a you know, a lot of things I cared a lot about.

You know, I was really excited about AI. I did think clinical practice would be really meaningful. At the time I was running a startup, a health AI startup that I was excited about. Um, you know, I have my family. I also really like to write and I I had, you know, this huge list of all this stuff I really like and I thought, okay, well, I'm going to have to

maybe whittle one of these things off the list. And that ended up being clinical practice. So, um, so yeah, that was how that came to be. And I I ended up steering into the AI space.

Yeah. And you you happy with that decision?

I am. Yeah, I am happy with it. I think it's a good match for my personality.

I mean, there were aspects of, you know, like I said, there were aspects of clinical care I really liked. There were also other aspects I had a hard time with. So, I mean, some people are really good at compartmentalizing. You know, some terrible thing will happen to a patient and they're able to kind of leave that at work and box it off. And I was not one of those people. So, I think it suits me to be, you know, a little bit further removed from uh kind of daily tragedy, if you will. I I'm still very passionate about helping patients, but I think for me it works better to do that from the tech angle.

Yeah. Interesting. So, my uh I'll drop this in now.

My daughter um she graduated from medical school a couple of years ago. She's now a practicing doctor in Australia. Um but she kind of did things the other way around to you. She went and did a different degree in um I'm not I'm probably going to get this wrong so I'm not going to say what it was. It was another degree and then she came in and and sort of got her head square about where she wanted to be uh and then decided to become a doctor sort of so she's tried something else and then became a medical doctor um and uh I think she's happy doing it. I mean she certainly tells me some of the challenges that she has and she says it's hard work very hard work being a doctor.

Yeah, definitely hard work and and rewarding. I I think it's a very it's a nice profession and my husband, he is uh practicing and so I still get, you know, a little window into that side of things through him.

All right. Well, cool. Cool. So, we got some good background. So, Glassbox Medicine, tell us a little bit about that aspect of your business. What What's that about?

So, Glassbox Medicine um originally started out as a blog. It still is a blog. It's glassbox. Com. I write about AI and healthcare there. And over time, you know, I was adding more and more articles, getting more and more readers. Currently up to 700, 000 lifetime readers across 185 countries.

Uh so there's only a few countries missing and the countries that are missing are like countries without internet. So

wow.

So maybe maybe they won't uh end up reading it. But um anyway, I you know I was writing this blog and and I was working on my previous startup. And then after I wound down my startup, I um you had started dabbling consulting. I thought I really should amp up this consulting angle and began work on more and more projects and realized it was really exciting to work with different teams across both different tech companies and medical organizations on implementing AI.

Um also did some you know AI projects outside of healthcare and so uh ended up kind of merging my blog and my consulting under this glassbox medicine umbrella. You dropped in that very quickly. 700, 000 users on um PE users people that come to your blog and read your blog post

total. So yes, over the you know the lifetime of the blog that's the total. It might actually be close to 800, 000 by now, but um yeah, I think I've read about written about um maybe 70 or 80 posts and uh the blog's been up for

probably five or six years. So yeah, that's the total.

And uh it's it's interesting to see over time which posts continue to draw traffic and which ones I I wrote this post one time about explaining chest X-rays and what you can see in chest X-rays. And this is really relevant because a lot of people getting into health AI one of the first things they'll do is they'll build a chest X-ray classifier to identify abnormalities. And that post gets so much traffic and I never would have imagined when I wrote that post I never would have imagined that it would turn into one of the most popular posts on the blog. So,

so how Yeah. Interesting. I mean, how many years did you say you've been running it now?

Uh, probably about five or six now because time flies,

right? And how many articles and blogs would you have up there now?

There's around 70 somewhere in that ballpark. Okay. And it's a wide range. Some of them are pure AI. So, I have some up there, especially from when I was in the depths of my PhD where I was dissecting some AI research paper. That's what some of them are. Some of them are squarely in the healthcare side.

So I have a series of articles on um explaining different body systems and medical terminology which can be really helpful especially if you're working on uh something with text where there's a bunch of medical terms that it's useful to know what they are and then I have a lot of articles that at the intersection where it's about something that is both AI related and healthcare related for example automatically interpreting medical images with AI.

Yeah.

So it is all around AI and healthcare and we're going to put a link in the description of this podcast video so people can go and have a look at that because it sounds like there's you know this is a really interesting topic isn't it because the US government well the US market for healthcare has recognized there's an enormous amount of administrative costs there's a huge burden um I believe maybe you you do too that there's an opportunity for AI to really help take some of that waste away because waste means that money goes into the system and it doesn't come out in terms of health care. It doesn't come out to you know the providers.

It ends up being wasted in the system and I believe it's a third of a trillion dollars. It's about $265 billion is wasted in administration. And um you're you're talking about something specific around around X-rays and imaging, but also there's this other side of of AI which I can or which we can get into, but let's come back to your your X-ray um AI. So, to me, this sounds like you get an X-ray and you put it into the computer and the AI model runs and says, "Oh, there's a problem here, possible cancer or some other artifact. " Is that what you can do? Is that where your research has been?

So yes, I I have done a lot of research and a lot of consulting projects in the medical imaging interpretation space. And yes, that's basically how it works. So you would take a medical image, it could be an X-ray, which is 2D grayscale. It could be an ultrasound, which is 3D grayscale, or you know, a CT or an MRI. Really any kind of medical image, and you put it into the model, and there's a bunch of different things you can do with it. One of the most popular things is just figure out what's wrong with it. So that would be a classification task where you might say, "Oh, there's, you know, cancer or there's a lung nodule or the heart's enlarged," something like that.

Um, but you can also do other kinds of tasks like you can put boxes around different findings. So not only are you saying, "Oh, there's a lung nodule," but you put a box around it. And then you can do uh segmentation tasks where you're actually identifying every single pixel that's part of some particular abnormality. So you might say, "Oh, there's a lung nodule, and you know, here's all the little pixels that are part of that lung nodule. " And um it's one of the most successful areas in clinical health AI. I believe out of the FDA approved medical devices, roughly 70 or 80% of those are medical image interpretation systems.

And there are lots of ways it can directly have a positive impact on patients. Like there's companies that do um kind of early detection of emergencies in medical images. So they can flag, you know, hey, this person's medical image shows that they're having a stroke. You know, go treat them. And by shortening the time to treatment, that can really improve patient outcomes. So, it's a super exciting area. Um, and I will also just throw out there because this is one of my pet peeves. You cannot just chuck your medical images into ChachiBT. It cannot interpret medical images and damn it.

Often come across people who think that IT CAN AND THEY'RE LIKE, "OH, we want to, you know, deploy chatbt to interpret these medical images. " I'M LIKE, "NO, STOP. YOU'RE LIKE, YOU'RE GOING TO HURT PEOPLE. " SO, so all these medical image systems are, you know, they're typically purpose-built for a specific task. Like this is you know this is going to be our medical imaging system for interpreting you know this particular kind of X-ray or something.

Yep. Yep. 100%. So that's how um me and my company got into AI was through uh vision systems and classification is something we did a lot of and I call that deterministic. I think you do too where you're determining a particular class.

We were determining people and traffic cars and bikes and things and we we call that a class. A car is a class, a bike is a class. And so it's determining with a certain level of confidence what that class is. Now we talked about that in the medical imaging space diagnosing a cancer. It's a class, right? So that's a cancer class and there could be other classes, stroke classes and things like that. But then you talked about something else which I didn't quite get. It was you called it segmentation, I think. What was that exactly? Can you just go over that again?

Yeah. So segmentation is where you get an image and you want to identify exactly where something is.

This can be especially useful if you care about how big it is. And um the way you do that in in terms of building it in AI is you actually have essentially a classification task but on every single pixel of the entire image. So if you feed it an X-ray, you're doing basically classification on each pixel and you want to know which class is each pixel part of. And then you can make these pretty visualizations where you you have your original image and then overlaid on top of it you kind of have traced exactly the borders of something and filled it in. So you know oh this is part of the you know this is all part of the lung nodule or you maybe you're doing it for whole organs or something.

So you know like oh this is all the liver and um what's cool about that is that if you have process your data so that a single pixel corresponds to a certain distance in the physical world then you can track the size of something over time. Uh, so maybe if you're looking at lung nodules, you contract, did this lung nodule get bigger?

Oh, gotcha. So, yeah, it's it's not obvious to I think people who are outside the space that when you look at an X-ray, it's not like looking at the, you know, it's not like having a um the liver in your hands where it's obvious that's the border of the liver. It's a little bit more nuanced than that, right?

It's not always 100% clear what's the liver and what's the not liver. And so this is where AI can really come in and and and really help the physicians. Oh, this is part of the liver and the cancer is either in the liver or out of the liver and that informs them about the diagnosis, informs them about possible surgery and all those other things. So I mean yes, there's definitely within medicine, you know, the the radiologists are they're the experts on all the medical images. So, you know, they'll tell you, you know, this is exactly where all the different parts are and this is, you know, this is what this rare finding is.

But then if you're not a radiologist, it can be difficult sometimes to understand exactly what you're looking at. And that is one of the possible deployment areas of these kinds of systems is helping physicians who don't have that deep expertise in specifically medical imaging understand exactly what they're looking at.

So should a radiologist be worried that they're not going to have a job in the future?

People have been talking about that for a long time. When I first started my PhD, I definitely bought into that angle. I thought, you know, in a few years we're going to have completely automated radiology.

I no longer believe that we are imminently completely replacing radiology because of a couple things. One is that um there's just so many thousands of possible conditions and the distribution it just has a very very long tail. So there's all of these super rare types of abnormalities and you might not see very many examples of them. And that's the kind of thing where a person tends to be a lot better at identifying those than an AI system. And so that's one thing where um where humans are still doing I think better than AI in a lot of cases.

And you know part of that could be a data limitation but um it is also partly due to the fact that a lot of these models do preferentially tend to do best on the most common categories. And then another reason is that um there's a lot of quirks in medical imaging that make it difficult to automate that might not be immediately obvious. So one example is that all of the different scanners can produce images that have differences which to a human are not obvious but that a machine can pick up on. So, for example, there might be some kind of processing artifact that's kind of distributed across the pixels that identifies like, oh, this image came from this scanner.

And that, you know, on the surface seems like, well, why is that even relevant to interpreting medical images? It's relevant because if you have a training data set that comes from a couple different places and the disease distribution is different at these different places and then the model can also see where they came from, it could look like the model's getting good performance, but it's not actually looking at the abnormalities. It's just looking at where it came from. So, there was a study back um I think it was back all the way in 2018 where there was a model that could predict pneumonia from X-rays and the performance looked really good.

If you looked at the numbers it, you know, they were impressive numbers, but when they delved deeper and did some explanation methods on it to see which parts of the image the model was looking at, it was looking at all this weird stuff. It was looking outside the patient's body and like the air around the patient's body. It was looking at the metal tokens that get put on the patient to say which side's the right, which side's the left. And it turned out that there was one hospital that had a much higher prevalence of pneumonia and a certain X-ray machine and then another hospital that had a much lower prevalence of pneumonia and a different X-ray machine.

And so the model had just learned how to distinguish what machine took the X-ray and that was all

it wasn't even looking up the patient's lungs. So that's something where you know subtle things like that can um can make it a lot harder to automate than it might seem on the surface.

Yeah, that's really interesting. So when you get down into the nuances, it's not as straightforward as throwing the image in and getting a result out. Um, and it's something so critical too. I mean, there's con there's real consequences and AI doesn't necessarily care about the consequences. I mean, you know as well as I do that especially LLMs are there to please you right in the moment.

Please you in the moment. It doesn't really care about its reputation as a doctor or as a radiologist which is something very human um and hard to replace. I mean we have accountability as humans and so the long term matters to us. We associate ourselves with a network of people and we want to be a good radiologist but I still think that there's you touched on a few other things. So perhaps the combination of AI and radiologists gives some better outcomes when they're matched together. Well, there definitely are use cases like that. So um the one I mentioned before with kind of early detection of emergencies.

I think that's one of the one of a great example of something where humans and radiologists are are both together. And what the AI adds there is that the AI can run very quickly, you know, basically immediately on images as they're flowing in. And so what it can do is you know you get this huge queue of all of these images and if the only thing you can see in your workflow is just kind of the you know patient's name or whatever you have no idea which ones you should look at first versus not. And so by having this AI system that runs behind the scenes and goes through all these images identifying things that might be emergencies.

It can fix that order and prioritize them and say hey let's you know we'll flag these three these have some alarming thing in them human. You need to go look at this now to just check if there's really an emergency here. So that that's I think a great example where you know you have an AI, you have a person, they're both working together as part of the same workflow and together they do better than either one by itself. [snorts]

So just to this manufacturer scenario to make it clear, somebody comes in from a car accident, they're perhaps they're unconscious, they're pretty bad in pretty bad shape. They can do their basic checks.

You know, the doctor will look for some things and then they'll go to get some X-rays, right? And there might be multiple X-rays. In a traditional setting without AI, they would call the radiologist down and there's this time lag as you get the X-ray, it gets printed, he does the work. All all the while this guy's unconscious on the table and they're waiting to know what to do. So, they're just in a holding pattern. And we can shorten that time. Is that right? Shorten that time that the person's in this no man's land of not knowing what to do.

Yeah, it can shorten time to to diagnosing so that it's clear earlier what to do.

And it can also catch problems that you may not even have expected. So, you know, maybe someone came in because they were having, you know, a really severe headache or something, but then they get some imaging and you discover like, oh, actually there's this other process going on at the same time. So there's there's an additional opportunity too where sometimes it'll uh sort of catch surprises.

Yeah. Well, this is great. Thanks. We got a good clear picture of that. So medical imaging isn't just X-rays. There was MRIs, CAT scans, those sorts of things. Is that also can it also be done with AI?

Yeah, definitely.

And um I've worked on projects on X-rays, MRIs, CAT scans, uh the kind of whole gamut and all of those um the I mean the main difference is that CTS and MRIs are just bigger which you know maybe years ago would have been a problem but now GPUs are so massive that it's really not a problem anymore. There's just these enormous GPUs that are completely fine training on these gigantic images.

All right. So we're familiar all of us are familiar with using the chat and so and this is done with the browser and we we talk to a model which is in the cloud and everybody has access to it with these types of models. Are they usually running in the hospital?

Is there a is there a cloud interface to that special model? Are they running on a local computer? Where where are the models actually housed?

That's a good question. I mean, for most startups in medical imaging, I believe those are typically cloud-based. So, they're running their own kind of server in the cloud, and that's where the image goes, and that's where the model is, and it gets interpreted, and then it kind of sends the result out. But there's no inherent reason why you couldn't have one that was locally deployed in a health system.

And I'm sure that there are some health systems where maybe they have an in-house model that they developed with their own team and then they could run that locally. So um yeah, in terms of infrastructure, you basically once the model's already trained, you basically just need a GPU that can run it or you know maybe a few GPUs depending [clears throat] on how big your model is and then you can uh go.

Has this mainly been startups doing this type of work or have we seen some of the big corporations as well coming into this and offering sort of that type of service? It's so within healthcare AI um I think the majority of the spending does go to startups.

I think somewhere around 70 to 80% goes to startups but there are certainly huge companies that work in the health AI space. So you know for example Microsoft has a lot of different health AI kind of projects that it does and Google has a bunch of health AI projects that it does. I think actually all to some extent all the big companies I mean um Amazon has its one medical and I'm sure they've got some AI associated with that. So yeah it's uh it's interesting because you know the big companies also have so much else that they're doing.

So if you're looking for a company that's purely focused on health AI that's more likely to be a startup because the startup's more likely to have a narrower mission.

Right. I was going to go down that path of well do you have u people that analyze X-rays for livers like is a startup that narrow or how does a startup find its own niche? What defines its niche? Can you give me an example?

Yeah. So um I had a colleague actually who started a medical imaging company and their initial niche was looking at catching lung blood clots in CTS. So the CT scan might have been acquired for some other reason or it might have been acquired because of concern for blood clot.

It didn't matter regardless of why the CT was acquired. If it had the patient's chest in it, it went into their pipeline and then they detected if there was a clot in the blood vessels in the lung. And that's really useful because um these clots can be extremely dangerous. And so if there is one, you want to know that it's there. And you especially don't want to miss it.

And the funny thing that happens when people interpret images is if you know especially for these kind of incidental findings if you get an image for some other reason um then when the radiologist is looking at it they're not primed to be looking for a blood clot and you know part of their workflow is to check for other things that they're not expecting that is like a normal part of radiology practice but um it's extra helpful in those situations to have an AI system that can catch this. And so the way their product worked is if they didn't see a blood clot, there would be a phone call that would automatically go to the radiology team and say, you know, hey, there's this image.

It has a blood clot that might have been missed. Go back and check the image. Um, and that was their initial product. So yeah, it's definitely possible that companies will start out with an extremely specific uh problem that they're solving. And that's, you know, for a startup probably better, you know, if you're starting with something that's highly specific, that that's better than trying to kind of solve solve all of radiology at once. So, are there hundreds? Are there thousands of of of medical imaging model providers out there now?

Uh, that's a good question. I I don't actually know. I don't actually know the total number of medical imaging startups.

I

Well, I guess the numbers are important, but it's just I'm thinking to myself, if I'm a hospital or a doctor or whatever, I'm going to have to subscribe to all these different um companies for all these different things. Like it's blood clots in in CT scans, it's uh it's liver damage from drinking, it's like whatever it happens. Do I have to have hundreds of subscriptions and and then push it out to all of them and or how do I know which one to use and how do I subscribe to them all?

Well, I mean, I think the interesting thing is that they're all solving somewhat different types of problems and the existing, you know, radiologists still very much have jobs and work within these health systems. And so, the baseline is that, you know, these images are getting interpreted and the radiologists are are um, you know, going through the images and interpreting them and doing that critical piece of the workflow. And so, it's not so much a question of like the images aren't going to get interpreted if we don't have these AI systems to do it.

It's more that, you know, in this particular setting, like maybe in this particular health system, could we benefit from adding a system that will solve a particular problem? And going back to kind of the lung blood clot example, you might be interested in that if you were a health system where you have a CT scanner and you do a lot of CTS. You're probably not going to be interested in that if you're like some random outpatient practice and maybe you send patients out somewhere else to get CTS, but you don't actually do them yourself.

So there's different kind of

there's different niches and uh and different reasons that an organization might want or not want to purchase access to a certain model.

So in the past we would have had the radiologist receiving the films and he would be the the resident radiologist for the clinic or for the hospital and he would be doing there might be multiple but let's talk about one and he would do it bas basically himself. Now he he's the guy that takes them and says okay let's try this model and this service and that service and push them out and get those results as well as his own interpretation.

So so actually I don't think in a lot of cases the the AI is kind of between the radiologist and the image in that manner. It's more like the AI is a parallel system running behind the scenes at the same time as the radiologist. So the AI system just is plugged in computationally and it gets the flow of data as it's you know as these images are being collected and then it runs on whatever you know parts of the data are relevant and then it will do different things with its predictions depending on what the purpose of the system is. So the radiologist will still uh do their normal interpretation of the images and they um don't typically choose which model it goes to.

It's more that like if a health system is has pardoned with a certain company then those companies models will kind of parallelize with the radiologist and then maybe the there'll be a kind of alerts that go out to the radiologist depending on what the models find.

Okay. So this this other company which is bringing all these models they're like an aggregator for all the models coming in from all the startups. And so they become the machines send those images to the aggregator. The aggregator says, "Okay, these are the models we're choosing from. " And then gives the results back.

The the hospital um system itself will typically have like a pack system and they'll have, you know, their images that go basically some computer somewhere um some computer somewhere which like in the old days only the people would be reading off of the computer. Then with you know if you have an integration with a some company that does medical image interpretation those images could also go to that company as well.

Oh right all right.

Well we went deep on the the process because I I'm trying to think if there's people can relate to it by what they experience when they have services and it does it sounds like it's all happening in the background and the main benefits are they might get a better assurances faster um and maybe in some cases even cheaper services. So if I was a startup, how would I go about doing such a thing? Do I need to get thousands and thousands of images and and throw them into some model training and then uh and then have I guess radiologists and other people guide that to um to to some result? Is that how would a startup would get going in this space?

Yeah.

So you need the thousands and thousands of images and uh and getting the data and making sure you know it's the right data is uh is often tricky and um I actually have some um colleagues who started a company to address that problem. So uh it's called gradient health and they actually aggregate medical imaging data for the purpose of then licensing it to companies that want to develop these models because a lot of times getting the data is really really tough. You know, if you're just trying to start a company and you're at square one and you don't already have anything, it's really really hard to convince a health system to give you their data.

They don't want to do that because they don't know who you are. They don't trust you. It's patient data. There's all these laws around it. So, um, so yeah, getting the data can be tricky. And you have to also be really careful about how you construct your data set. So, it does need to be large enough. That's one of the biggest barriers is I I work on a lot of medical imaging projects and sometimes people will come to me and they'll say, you know, oh, I have 2, 000 images and I really want to get this to be almost perfect performance.

And I'm like, well, you know, it's probably not going to be almost perfect performance with only 2, 000 images because you got to the, you know, the higher performance that you need, the more training data you need for the model to be able to perform well. And then also the data needs to be reflective of where you intend to deploy this.

So something that can happen um that can sort of derail things is let's say you create a data set and you create it entirely based on this one hospital system and you used you know all this data that they had and you train your model and then you want to go deploy it in some other hospital system you're going to have to be really careful when you do that um and and do a lot of work to check how it performs there because there's a lot of reasons why it might catastrophically fail.

So for example, if it has learned any of these correlations between like different scanners in different departments correlating with different diseases, anything like that and you try to deploy it somewhere else, um there's just a lot of ways it can go wrong and you want to um make sure that the data is is appropriate and and if the data distribution changes over time, maybe you have to retrain your model over time as well.

Yeah. Yeah. Well, that was fascinating. We took a real bend in the river there to go and explore this world of medical imaging in AI.

Um, and we can I think we could keep talking all day, but um, let's talk a different angle on this because you've been working in AR in other areas too, right? [snorts]

Yes, I do work in other areas. So, some uh, projects that are totally unrelated to healthcare and then also within healthcare, other sides of healthcare. So, you know, billing, OMIX, EHR data, notes, lots of different other parts of healthare.

Let's talk about billing because this is an area we've been working in as well. So tell us what you've been doing there.

Uh so I did work on a project related to insurance claims and claims denials uh and trying to do some automated processing there and that I mean that's a huge area just in terms of the broader you know health AI landscaped revenue cycle management automation is is huge.

Yeah. So, back to the top level, the billing side of things seems so complicated to me. Coming from Australia, um, which seems to have a very simple system. I think they've got half a dozen insurers and over here you got hundreds. You always have this complex arrangement at the moment where the physician has to justify the coding so they can get the rebate. Um it's much simpler in Australia than it is over here.

And I believe and we're sort of stepping into this ourselves is that AI has the benefit is it can take all of the documents that are associated with that particular code or sequence of codes that you think it's the and then read them all and say well yeah this one matches best to this one by by by its ability to read fast and match. Is that your understanding of one of the benefits we have coming with AI?

Yes. So, it does help simplify the enormous unnecessary complexity of the uh US kind of insurance paperwork thing that's going on. I I recently read um The Healing of America, which I think was a New York Times bestseller.

Super interesting look at the US healthare system and healthcare systems around the world diving into all these ways that there's you know many other countries that have solved certain things which remain a problem in the US one of those being the very complicated nature of the insurance reimbursement so yes I think this is also an example of where you know the ideal solution to this problem is just a policy change like that's the real ideal solution is we just dramatically simplify the way that billing works from like a legal level BUT YOU KNOW IF THAT doesn't happen, then AI can certainly help by going through all of the different painful steps of that process and trying to save a person

from having to manually deal with them all.

Yeah, I heard us one of the guys on the podcast earlier this week said that a fourperson team would take an hour and a half to do this work. Now with AI, they're doing it in one hour and a half. Right? So instead of set of so four person team, half a week down to an hour and a half. So that's like a 15 to 20 um improvement in in uh in time as well as being quite a tedious process. [sighs and gasps]

Yes, it's definitely definitely a great opportunity. And I think also um something nice about it is my understanding is sometimes their business model is like a percentage of extra money that they're getting you or something like that.

So, you know, you may end up with more reimbursement than you otherwise would have. Especially, I think in some cases if you're, you know, independent practice, maybe you lose out on possible revenue just because you don't have the personnel to go follow up with every single denied claim and all of this. And so, if the AI system is there doing that for you, you might end up making more money.

Yeah.

I was well only a few weeks ago a dental practice opthalm eye an eye practice an opthalmologist optometrist one of those um said that they were fairly decent sized practice they said there's probably $20, 000 a year still left in the system we we we just can't find it we we we just kind of write it off and I thought to myself $20, 000 that's a nice holiday to Europe for uh the owner and his wife I mean that's got to have some value hasn't

Yeah. Yeah. I mean, it's and the AI system can certainly unlock that because you if the alternative is that you're spending like a month doing all this paperwork, then the $20, 000 uh becomes possibly not achievable.

You know, maybe you have no extra staff who can do that month.

Yeah. Yeah. 100%. I mean, our direct example is the Australian physical therapy clinic that's saving $2, 000 a month because the AI is stepping in and doing some of the heavy lifting. Now, it's not perfect. We have quite a few that fall outside of what AI can handle for various reasons. I mean, one of the problems is that all of the different insurance companies have different ways they want you to interact with them. Uh, there's no standardized way of sending the invoices through. So, some of them are all batched up. Some of them are you have to go in through a one-time password type portal and that slows everything down.

And we've kind of got ways to speed that whole thing up. And then what we found was the presentation of the data by collecting all those pieces and showing the bookkeepers and the accountants. Oh, this is what we're seeing the three panels. Okay, I know that that number isn't the same as that one, but I can tell by the rest of it that it matches. So, that's that. Go ahead and reconcile it. Um, we're able to get sort of 66% at that. Our goal is 80%. Um, but we're here at 66. I think we started at 10. Now we're really starting to get somewhere with it.

Yeah. I mean, that sounds like a big difference for sure.

Yeah. Yeah.

So, you're working in the same space in a much more complex area over here in the US. It's not only is it oh gosh 20 times more people here maybe 15 times more people here it's also fragmented through states much more than it is unified in Australia there's an Australian federal system of course there's a state system too we got so many more insurers and you got the billing codes on top of that um it's really a challenge isn't it to unravel the complexity now with AI it can actually make things worse so I had a meeting with someone this morning and they were trying to drive AI into their organization and she was frustrated.

She said, "Look, the the the founder wants to use AI and he's making me do this, but it's upsetting my workflows and I can't use it sometimes and I can't. " And she's on the sales side of things. Um, the problem with complexity is that it can make that worse, right? It can make put another layer on top and make things even more difficult. Is that what you're seeing too with this attempt to put AI in healthcare?

It definitely can happen. And I think that that the issue you're describing is a common issue where um it's basically choosing the wrong problem to solve.

And I mean what I mean by that is, you know, maybe you're saying, I really want to use AI to solve this problem, but there's a lot of reasons that might not be a good decision. First of all, maybe it's not a real problem. Maybe it's actually just a minor annoyance and there's no reason to put a huge amount of money into trying to get AI to do it. Or, you know, maybe it is a real problem, but AI is not the right solution. Maybe what you actually need is a policy change or maybe what you need is some change in your interface or something else. Like maybe there's some other way to solve it that's not AI.

And then um you know, I could go on, but like basically if you've picked the wrong problem to solve, you're trying to force AI onto it, that can definitely end poorly. Uh because you need to understand like who's going to benefit from this and why are they going to benefit and also why specifically is AI the right tool to solve this particular problem.

That's right because to at the end of the day I I see AI as fundamentally a pattern recognition tool at the moment. Um LLMs are big and they have other sort of ways of making that work better than just basic pattern recognition.

But if you can't sort of stabilize the data in and what you're asking it to do, uh you're not going to get good data out, right? Um that it's kind of this black box. The way I see it in simple terms, you got to get a clean data in. So to me, there's always a good analysis section that has to happen or discovery element to what's really going on. And the danger is saying, "Oh, we got to get AI into the business. The competition's coming. And I got to do it and let's just ram AI in here and see where it goes without doing this analysis piece at the start. As somebody else said to me the other day, everybody's an AI expert, but almost everybody's getting it wrong or something like that.

And there I mean there is a lot of material about what AI can do that may not be entirely accurate. I mean, something I've come across lately is, you know, it it does change over time. You depending, you know, there's like a kind of flavor of the month for what new thing people think AI can magically do for them. And I think something I've come across a lot recently is a lot of the way that AI agents have been framed is that they're basically people. And, you know, you just have this AI agent, you just tell it what you want to do and it just goes and does it. And they're very much not like people.

You know, process of setting up an agent to do something for you automatically is very different than hiring a human being and training them to do that task. And I I feel like I end up having all these conversations where I have to disappoint people because they're like, oh, you know, I want to have an AI agent who's going to completely replace every single thing that my front desk receptionist does. I'm like, well, that's going to be really tough BECAUSE I'M SURE THAT THAT PERSON IS ACTUALLY DOING HUNDREDS of different things. And if you wanted to have those automated, you're going to have to look at every single one of them separately.

And you're going to have this deep integration with what with whatever software they're using. And maybe they're actually using four or five different pieces of software. And then you need integrations with all of those. D like the agent can't just automatically integrate itself. So um yeah, it's it's a it's an interesting time and you know, I'm sure six months from now it'll be something else.

Yeah. Yeah. Look, um, Rachel, we have had I've had a lot of fun here and I really want you to come back, but we're running out of time. Um, is there anything you would like to say or you've got any questions for me?

Um, what are you most excited about in the AI space these days?

[sighs and gasps]

Oh, what am I most excited about? Um gosh, you really put me on the spot here because there's there's lots of things that I'm really excited about. Um I'm not going to say data centers in space. I don't think that really excites me. Um

it does sound like science fiction though.

Yes, it doesn't. I guess that's that's a part of the luster, but doesn't really excite me. You know what excites me? And it always has been the case with technology. I find out something that I can do so much better with a piece of technology and I get this a oh wow moment and AI's been doing that for me a lot lately. Uh and use a non AI example when I first had my first computer.

Can you believe it was a time before computers? I was a terrible writer. Still am. I could use a word processor and then I could fix up all my mess, you know, and that was a wow moment. And then I had Lotus 123 which was an Excel spreadsheet. That was a wow moment. And then with the web coming on where I could get access to company websites, that was a wow moment. And I'm getting a lot more of those. They're coming faster and faster. And so my day gets better. I used to have to spend a lot of time with people trying to get uh reports done or presentations or understanding my research. I can do it so much quickly.

It's cut my time down and that makes it more fun because I can make progress really quickly on my own. So that's what excites me. Thanks for asking.

Yeah, it is a definitely there have I've also had a lot of wow moments recently with what AI can do.

Yeah. Do you know what a 911?

You know, I think the first time I chatted with a large language model that was actually fluent, that was a wow moment because in graduate school when I you was diving into language models, they were still pretty bad. So they would do sentences like, you know, the cat walks walks walks. They're just stuck in these loops and or they would have gibberish or things like that.

And so I think the first time I had a a chat with the language model where it actually, you know, it was a fluent conversation. It all made sense. There weren't any repeated words or, you know, it didn't start going off the rails about something. I think that that was definitely a wow moment for me.

Yeah, I can understand that. So, I have one more question for you, Rachel, before we wrap up. Uh, AI, is it hype or help?

I'm going to give a difficult answer to this, which is that it depends. So I mean some AI is hype. So I mean anything that's claiming to solve all your life problems with AI that's that's definitely hype in you know more specific example in the healthcare setting.

Anything that's saying AI is going to completely replace your doctor or completely replace your nurse or something that's also in the hype bucket. Um but also there's a lot of AI that is truly helpful and it's having a really positive impact now and there's also new forms of AI that are being developed that are going to become helpful soon. So it's it's both.

Yeah. Well fantastic. That's a great answer. I've really enjoyed having you on. There's so much more I want to talk with you, Rachel, about maybe we can set up another time for you to come back. Um, also stick around when uh when we finish up here because I want to have a quick chat to you. All right. Well, bye.

Sounds great.

Thank you.

Well, I hope you enjoyed that as much as I did. Rachel was fascinating to listen to. We did a deep dive into medical imaging, what's happening behind the scenes in the medical industry, and how it's being revolutionized with AI. So when you go to the doctors and you get an X-ray, you might not know what happens behind the scenes. Well, traditionally that has been a radiologist who has manually looked over and made their recommendations and reports back to the physician. Well, now AI is in the picture and it's doing a really good job of supporting, not replacing, but supporting radiologists.

There's also lots of startups out there who are creating unique models solving particular problems around healthcare. For example, is it detecting um clotted uh arteries in the heart or is it looking at some liver deformality? So there's many startups in this whole ecosystem with specialist models that they're creating. So how it works is the hospital system connects through of and often an aggregator and that through an API that information comes back the image goes out the information comes back uh with a diagnosis or with a report associated with that image. So that then comes back to the radiologist and it supports and helps them.

But we also found its limitations not very good and she talked about an example of it was b it was biasing the results because there was a different machine being used in one hospital than another and that was swaying and skewing the results. So look there's plenty of ways this can go wrong. Uh and we talked about a few of those but overall the trajectory is very positive. It's making um patient care better. And one fundamental way it's doing that is making that diagnosis quicker. And we talked about an example of something coming somebody coming in to the emergency after an accident, a car accident. Maybe they're unconscious. They go get some scans.

They can very quickly get the information back and often see things that they might not have been looking for so they can get the treatment going very, very quickly. Well, I hope you like that show. If you know more about Rachel, follow the links in the description. If you want to know more about Skillion, then please do the same.