Can You Trust Clinical AI When No One Is Asking the Right Questions?

Chris Hutchins: Welcome back to The Signal Room. I'm Chris Hutchins. Quick thing before we start, I wrote a book. It's called Beneath the Signal. And it's about the human work behind trusted data and responsible AI in healthcare. Search Beneath the Signal on Amazon now and pick up your copy. Okay, here's our episode. Welcome back to The Signal Room. I'm Chris Hutchins. Almost every conversation about clinical AI is about the model. Is it accurate? Is it trained on enough data? Can it pass the exam? My guest today argues the model was never the missing piece. Declan Kelly trained in practice as an NHS doctor before founding Eolas Medical to fix a problem he lived firsthand. The right clinical knowledge, not reaching the bedside fast enough. Eolas is now live in about 500 sites, including roughly 90% of acute NHS trusts, plus systems like Stanford, Mass General and Boston Children's. This is a conversation about the layer clinical AI keeps skipping. Not the model, but the hospital's own governed knowledge. Declan, welcome to The Signal Room.

Declan Kelly: Thanks very much, Chris. Really good to be on. Thanks for having me.

Chris Hutchins: It's my pleasure. Well, before we dig in, I'd like to hear a little bit about your backstory. So I mean, what's your why? How did how did you end up in medicine? And then probably more interesting and exciting for me is how did you end up building a technology at Eolas? Because I think I'm hearing a lot about the inadequate systems and methodologies that are being used in AI, but data foundations and stuff too from clinical people who are tired of us hearing we throw around the word operability, but we've never interoperability, but we haven't really done much with it yet.

Declan Kelly: Yeah, yeah, no, a lot to dive in there. And yeah, I can jump in with a bit of background on myself. So I'm I'm Declan, I'm a medical doctor by background. I worked clinically as a physician for about five years, mostly in emergency medicine, and absolutely loved it. I was very much hurling down that path and loved treating patients, loved solving problems. And I suppose two things really happened. You know, one was a push and the other was really a pull. And and the push is that I think there's no health system in the world that's immune to this. We're facing you know exponentially increasing pressures across every area of healthcare in general. There's never enough doctors, nurses, pharmacists, clinicians, and that doesn't scale. And a lot of systems are actually at breaking point because of that. And that's kind of one side. The other side is that technology is rapidly evolving over the last over the last 10, 15, 20 years, and the potential there to really impact healthcare at a meaningful level is is really, really exciting. And I think both of those things combined led me down the path of launching Eolas. And Eolas is essentially it's the Irish word for knowledge or or information. And when myself and my colleagues were treating patients in our own health system, what we found it was very difficult to access the knowledge of our own health system. So someone's having a major heart attack, someone's failed their cancer treatment. In this specific health system, what do we do next? You know, not the external knowledge, not what's the latest best evidence, that is really, really important, but actually practically what can we do next? You know, what's on our formulary, what drug should we give, what's our policies, procedures, guidelines, etc. Usually that information either didn't exist or existed on SharePoint or some outdated system that doesn't interact at the point of care. So we developed Eolas, it uses AI to index all of that content for the hospital, and then the clinician can ask a question, get an answer within seconds. We launched that about four years ago, it's live now in about 500 sites across the world. So we're now up to 90% of all acute NHS trusts use it, and about 500 sites, you know, Stanford, Mass General, Boston Children's, a number of other sites are now using this as well. So it's really exciting as supposed to be building and combining that you know physician element with the technical element.

Chris Hutchins: That's phenomenal. I I've been on the data curation end of it for a long time, you know, for various roles, but you know, data and analytics, but also running IT organizations. You know, interesting enough for about 15 years at Mass General Brigham as well. But the the challenges tend to be the same ones over and over and over again. And you know, the the competitive landscape when it comes to the technologies that are being used. I just think that we've we did what we've done in almost every kind of technology space when there's an opportunity to improve something when all of a sudden someone says, okay, we can do the electronic health record. Well, everybody decided to get into it, and then we've created this unbelievably ugly spider web of systems that some play in the exact same space, some play in adjacent spaces. All of them go at things from a different lens, and they're they have differentiators that they believe are differentiators, but it all creates challenges on every level, not the least of which is what clinicians are dealing with. So yeah, hopefully we can get the accurate building that everyone thought we could get. But yeah, in the meantime, we've disrupted workflows, we've just disrupted the patient encounter, the things that you decided to go into medicine or to do. We've challenged that. We've created these wonderful terms like pajama time that I heard about not too long ago for the first time. I was mortified. I don't think that should be a thing, by the way. So I am a strong believer in eliminating that term for a vocabulary because we're giving back time. So that that's that's what my my number one priority is, and the the the lens I'm gonna put on when I'm helping people to determine where you're gonna put your time and energy in AI. We gotta start AI, I think. I I want to kind of get to the what what the pivotal moment was too. So you you're why, obviously, there's a there's a passion for taking care of people, helping them to get better. When you were trained and practice, practicing for a period of time, what was the moment that you realized that the missing piece was not the technology and it wasn't even a better AI model, it was really the the institution's own knowledge.

Declan Kelly: Yeah, no, it's a great question. So I suppose there was a few things in combination that that that led to that point as well. So I I I obviously obviously as a back I have a background as a physician, but I also have a background as a software engineer. So I started developing solutions for our own hospital because of that. And it was really I started with the problem that I had. I tried to solve a problem that I had and my colleagues had, and it was accessing that internal information, that internal knowledge. Generally accessing external knowledge isn't very difficult. You know, you have excellent platforms like you know, before AI, excellent platforms like UpToDate, etc. And you're able to access that knowledge. You have national level guidelines. And generally, yeah, there maybe used to be a couple of searches away, but you're still able to access them. Now that's you know infinitely more accessible with the implementation of AI. So you use OpenEvidence or Doximity's models or UpToDate say and you can get external knowledge immediately. But that only takes you so far, you know, it might get you 50% of the way there, but that kind of last mile of the decision of you know how exactly do I refer them to the cardiologist or how do I book this CAT scan when no one is in the hospital and and and and I'm a resident doctor, etc., all of those decisions just live in people's heads. And and that's when it comes really, really difficult to get that. I would say that the the technology actually has played an enormous role in unlocking that. That knowledge was always there. It was either codified in policies or lives in people's heads, but the technology now has enabled us to unlock that and get the answer within seconds. So we we actually built the platform before generative AI, and we were using normal natural language processing techniques to try to extract that knowledge, try to organize it in using graph databases, etc. And the usability just wasn't there. And there's one thing about having all the information, that's one of the bits of the problem that needs solved. But the other bit is that clinicians need to be able to access it within seconds, and if they can't access it within seconds, they're they're just not going to use it. So that is the thing that the models have unlocked, and that's the really true valuable pivotal point that that is exploded excess here. And downstream of that, the benefits to patients are are are enormous. So that's really a few of the pieces that came together.

Chris Hutchins: Yeah, it's a it's been a problem for for such a long time. And I I think the piece that has so been so clear to me as I've talked to clinicians for a long time, but in particular over the last year, it's is that we are we we we're just not really clear thinking sometimes around what is the what's the most important piece of information at any given time. And we don't understand that the risks and how how you guys are thinking and what the risks are that you're you're weighing at that point and you're having a conversation with somebody for the first time. And the question that always comes up to me is what don't I know about this person and is it important? And AI is not really going to solve that kind of a problem, especially if we're not even paying attention to the silos and the the data that we the missing data that we have. So I I you know I'm curious to to hear a little bit of your take on this too. But the the gaps in the data to me are the unintended and more more frightening levels of bias that we might be dealing with in healthcare. Talk a little bit about what your experience is is like in that for just for a moment. I know this is a little bit of a deeper thing, but I I think it's helpful for people to understand what you're evaluating as as a clinician and how big of a deal is that gap when you when you don't even know what it is.

Declan Kelly: Yeah, no, I think that's that's a that's a great question. There's multiple different angles to kind of come at this, and I'm a fan of analogies. Sometimes they're very good, sometimes they're not very good, but the analogy that immediately just came to mind here is that I think you know in any health system and and in any company in general in the world, there's lots of silos of different data. You know, you have you have the labs data, you have the packs data for all the radiology scans, you have the EHR data, you have lots of other data, and it's kind of all floating about, and nothing's talking to each other, nothing's there, and you and they're all floating about in this kind of sea. Ultimately generative AI has the potential to act as the gel around all of that and to understand and be able to query that in in in natural language and get you access to all of those things. But the system has to allow for that and it has to be designed in in such a way for that. And I think there's two separate problems you touched on there. So one is the data that exists that we can't access seamlessly, but then there also is the data that just doesn't exist and we or really bad, messy data that we need to capture in a different way. And AI is helping solve both of those. So for the data that does exist, the AI is acting as a connector. So I don't know if you know in your in your own team we use Claude or or ChatGPT, whatever it is, but even in our own team, that now connects every aspect of our company. It taps into you know the conversations that we have with customers, it taps into all of the licenses that we have, etc. Etc. So at any one time we can just ask a question and we get an answer. So there's this brain of our company that exists, and that's ultimately what we've built for health systems, this brain that exists that's able to tap into all of those and then bring and surface the answer to the right point of time. The other problem is okay, the data doesn't exist or it doesn't exist in the right way. AI is also helping solve that. So the ambient listeners, for example, are doing a brilliant job of solving that. You know, normally, say if you if you're the doctor and I'm the patient, we'd have a consultation and then eventually you will go and write your notes, and probably you're gonna write down 10% of actually what was said. Ambient listeners now record 100% of what was said, they transcribe it down, and and for the most part, you know, give or take the edge cases, most of that is captured in a very accurate way. That is an enormously powerful downstream because once you have that information, you can take that, you can autonomously query a brain of your organization to decide what's next, and you can probably start taking agentic actions to say, okay, right, let's organize the scan, let's organize the labs, etc. Etc. So that the power of being able to join all of that up is you know, I can't it can't be understated. And I think I actually think we're on a path now towards the first autonomous health system, and and it's going to take years and probably decades and a lot of stuff to figure out. But because of all of these pieces that are there, they're now being joined up by by this new by this new technology.

Chris Hutchins: Yeah, I the there's a there's an aspect to it that still from from my standpoint, it becomes a a bit of a challenge. And it it I don't know that we ever solve it so much as we're just mindful of it and we're we're doing everything that we can do to account for the the difficulty that exists because of you know people's belief systems, cultures, norms, things like that that make somebody willing to disclose certain things versus not. I'm sure I don't know how how it is in every country, but I know for sure in the US, once somebody turns 18, the the likelihood of a male, a healthy male, actually get being seen over the next decade or so, unless he's out of an accident of some kind, maybe a broken bone, you may not see them. And so the the being proactive and taking measures to maintain their health oftentimes there's just a big a big gap, and all of a sudden there's we're still seeing the same kind of problems manifest later because we just don't know what's going on with them to even do like you know, gaps in care that we tend to do for population health service type of arrangements that we have. Yeah, I I think I think that that the unintentional, unintended stuff is the piece that has always been a something that I've I've always hated that we don't have a way to be a little bit more proactive in thinking longer term from a research standpoint. Research doesn't tend to get put together with that long horizon that we want to follow somebody for 50 years to see if we can actually make a massive improvement in preventive medicine. But we can keep talking about it and keep trying to inspire more people, and maybe somebody will will find a way to fund it and get that done in the next lifetime or two. I I'd love to think it's happening right now, and I'll see an announcement in the next 10 years of my life. But yeah, maybe you know of one, but I I'm not sure I've I've heard of one yet.

Declan Kelly: Yeah, yeah, no, I think I think it's a great point. I think there are a lot of gaps there, and but I I I would be I suppose quietly optimistic about the direction of travel in general. I think the way the way I kind of view this is you know I think gen I had this conversation today actually with an ex-colleague, and healthcare in general, at least Western medicine, we we we it's a very disease-centric model, and it's it's not as powerful on on prevention and it's not as powerful on observation. We're starting to move in that direction, but it's very much once you get sick, you go to the hospital, and therefore you're there's a lot of data you're not capturing before you get sick to be able to predict that. I that's now slowly starting to change, and for a number of different reasons, the incentive now commercially for entities is to go out into the community and to get these patients before they actually get sick to actually and capture them and enable them. And I suppose one perfect example for that is ChatGPT. So, you know, over 50% of queries in ChatGPT are healthcare related, and they're very much making a play for for health. And they're now acting as almost top of funnel to funnel in patients to these health systems. And they're now being able to say, okay, you know, it sounds like these symptoms are dex a dah-da-da. This is where you can go and get these types of operations, this is where you can go and get these types of you know, follow-up scans, etc. So and that's now starting to follow people longitudinally. What else is now follow starting to follow people longitudinally is all of the observable health data. So their you know, their Apple Watch, their Whoop, et cetera, et cetera. So all of this kind of the world of like health and wellness is now slowly starting to merge with the world of you know actual healthcare. And that will start to become very, very powerful if that data is leveraged in the right way.

Chris Hutchins: Yeah, that that's I I I I love hearing this stuff. I think I I was talking with a with a physician just about a week ago, and I mean some of the things that she was working on, she she kind of found herself in a in an uncomfortable spot at one point because they were making such good progress with some of the technologies that they really wanted to push past some of the limits that she was comfortable with. So it certainly made it challenging. But the but the cool thing is that some of the breakthroughs that are coming are causing people to have the an excitement for sure, but more of a passion and a drive to push the envelope to to go further. So I I love hearing that there's progress in in in the preventive space because I I think we need a lot of a lot more help with that. I think I was just talking to somebody about the the marketing stuff that we've done over the last few decades and just how much we do, and I say we as a society, I don't mean people in healthcare, but that are driving people to be consumers of things that are probably not the best thing for our health. And you know, there's a constant evolution, it seems, in terms of human diseases. We I mean, how do you how at this point are we discovering new things? I think we're kind of incubating them by some bets, some maybe poor decisions that we do in terms of what we put into products that we decide to consume. But I don't agree. So I I I really want to kind of get into the the the part the the second hat that you wear, probably I don't know how many you have. I I suspect it's more than more than just a couple of them, but you recently closed a $12 million Series A, and you now live in roughly 90% of acute NHS trusts and systems like Mass General Brigham. What changes for you know operationally when you go from wearing your clinician's hat to being an executive and managing a company managing the the evolution of products that are being rolled out across multiple hospitals and organizations?

Declan Kelly: Yeah, it's it's a it's a great question with m multiple different multiple different answers. Right if you'd asked me at a different stages along the journey so far. I think I think very initially what changes and what needed to change was actually the mindset of what it is to practice medicine as a practicing physician and and the mindset of what it is to try to build a company. I think as physicians we're trained never to be wrong. It's where it's deeply uncomfortable for us to be wrong about things because there are serious consequences when we're wrong. Building a business and finding where you can add value and fix problems when you're starting from zero, you're probably wrong nine out of ten times, and then you have to find that one time where it's right and it's working, and then iterate on that, and then continue in that upward spiral spiral and that upward cycle. So that I suppose that that was one thing that definitely changed and that that that that's that sort of mindset. The other thing that then changes you have to rapidly be able to think at systems level. As an individual physician in the emergency room, you know, you see one patient at a time, or maybe you have five or six patients on the go at a time, but that's that's it. When you're now trying to scale across you know 500 plus sites, you have to you have to start to think in in systems, which is really important. And you also have to start to delegate and give control to people that that that you trusted that are, you know, are going to do the job well. You know, the saying of if you want to go fast, go alone, if you want to go far, you have to go bring along other people with you as such. So I think that's that's that's the other thing. And to be able to do that effectively, communicating a vision of what potentially this technology can do for healthcare is really important so that you're actually able to bring along the team with you so they can see actually, you know what, that is probably the direction of travel, and I want to be involved in building towards that. So there are there are a few of the things and a few of the skills that definitely you learn aspects of them as a physician, but you you you you're not able to put them all together until you have to put them in practice in a different in a different environment.

Chris Hutchins: I I've always found it amazing when I talk to a physician who has done software development. I I I don't I can I can't comprehend how you make space in your brain for that, which you have to train for so long. All the training education, then it's the hand-on training and you know staying up with all the journalists, you know, you know, j just the natural evolution of things to be someone practicing medicine and still having that the energy, first of all, but then the the mental capacity to really dig into how do you build systems and do so do certain things to deliver capability that has never existed before. I actually worked with a neurosurgeon, a pediatric neurosurgeon when I had some journal years ago, and I I was just blown away because the guy would like literally he'd go home on the weekends and he would like develop stuff in his software in his garage. And he'd come in and he'd have some great new feature for one of the systems he'd been building. I'm like, I I don't know. There's people functioning at a higher level than me. And if I sleep seven hours and I run into somebody like you, I start to think, geez, I probably should have stayed up a little bit longer and worked a little bit more. But but I it's it's just I I I really admire and I respect that you know that that that's just a way that you're wired and it carries well beyond even the patient care. It's just pushing the envelope from every direction that you can to to really fulfill that that passion that's that grows inside of you. Absolutely amazing. Because a lot of the focus right now in the technology is around the models and Whether they're effective, whether they're accurate, all of that. But I think for for from what I've been reading that that you're talking about, it's it's first really about what the hospital knows. But maybe talk a little bit about the difference and and how that really you know looks from a practical standpoint when you're working with an organization trying to to help them to implement you know AI. What's that practical difference between a general clinical model that knows things from published, you know, published literature versus what it needs to know about one specific hospital to be safe at the point of care?

Declan Kelly: Yeah, yeah. So I think I think really the example I always give here is that this is this technology is is no different really in how we should assess it and implement it compared to any other technology that that has been implemented in the system. And what do I mean by that? So I take an example of a troponin blood test for detecting a heart attack. So you know, you come in with chest pain in the emergency room, you do a blood test, usually done within three hours. And that blood test isn't a perfect test, but it's extremely useful because of all the systems that have been put in place around it. And that's exactly the same truth for these models. The that blood test, for example, if you do it at three hours after someone's chest pain, it's got roughly a 98% chance of picking up a heart attack. That means that 2% of the time you're going to miss a heart attack just using that blood test alone. So what the systems do is they put other questions around that, they put other tests around that, ECG, etc. Etc. And then they're able to build up a safety net and a system around that, and then they have the clinician's judgment. So you take something that has a you know on paper just as the model itself or as the or as the technology itself has a 2% error rate, and they've they've turned it into a part of a system. Implementing these models into clinical practice is exactly the same. You know, AI is is probabilistic by nature, so therefore it's never 100% accurate. And it will never be 100% accurate in and as of itself because it's it's a statistical model. And therefore you have to build the systems in and around it. And I think understanding how the core functionality works, work, and feel, and how much system you need to build in around it is is is the ultimate key. Really to the core of the question is is the difference of the model itself that's not trained on your own data, and then a model that can access your own data. I think the when these models first came out, there was a there was a query around do they need to be fine-tuned based on your own data? Do you need to build your own model yourself? They probably don't, and that's the general consensus at the minute, and that probably will be the consensus moving forward. So you should you you will be able to use a model that is very, very accurate and then build what is called a RAG-based system around that. So it's able to call on and pull those those connections in. And that's how that's how most of these are now all being built out. Now, life sciences is different. You probably have to find you have to fine-tune models, and you have to build your own models from scratch that are brilliant at that one use case for your own health system, been able to make sure that it has access to the right data, and then the model can call on that data when it needs. That's that's really the key. And that's all about setting up the infrastructure and setting up that system.

Chris Hutchins: Something that has been prevalent in in some of the earlier conversations that that I've been hearing is it kind of gets to the challenge of the public access to the capabilities that are going on. I know when OpenAI and Anthropic kind of pulled the trigger on on rolling out, it kind of put some urgency in the in some fire in the in the organizations across the US for sure, that were maybe not ready to go at the pace that they have to now. But one of the things that I've realized just thinking about my own experiences I can actually sit with across from a physician that knows me well, and there's just certain things that they're gonna be able to pick up on because that you know I've seen them for a long time. Talk a little bit about the context piece of it, if you wouldn't mind, and and what your perspective is there, and you know, how you advise people now, whether they're clinicians or even friends, neighbors, family, you know, patients, around how do you how should you think about it? Because I think that context piece of it's one that I wouldn't even know what's important to half the time. I'm gonna give you a real example. I take a couple of medications, I someday is going to remember the name of it, but I certainly won't remember the dosage. But if I'm you know using the AI myself and I'm asking it questions, what recommendations it gives is going to be based on a number of things, but certainly things that it would look for that you that it's not looking for, that you might look for if you're my doctor and you know me. I think that matters, right?

Declan Kelly: 100%. Context is key. Context ultimately at a systems level takes these models from being nice answer engines to actually being able to be implemented across the entire systems and then actually take decisions and become agentic. Without context, you we can't have agentic actions being taken and we can't build towards autonomy. And with it without proper context, these are nice answer engines, essentially. Now very, very expensive to train and very big models and very nice answer engines, but ultimately that's that's all they are. They're they're just they're just word predictors of the next of the next thing without the context. So the context is extremely important. And it's all about building the system around making sure that the model has the context. I use I again, I use us, for example, as a company. We our instance of Claude, for example, has the context of most of our data sources that it needs within our entire company. That makes that platform then extremely powerful for anyone coming and joining Eolas. And that's what health systems are now starting to work towards. The problem is that health systems are enormous and they have data from multiple different locations that is very difficult to interoperate. And that's the problem that they're facing. That's gonna be the difficulty here, the the ability to make sure that all of those are getting ingested and then accessible by the model. That's the model's not gonna be the difficult thing here at all. It's gonna be the context layer.

Chris Hutchins: Right. It kind of leads into this. You you're putting putting the emphasis in a in the this knowledge layer, I think, that it's it's an important thing to think about in terms of what's how it's different now because of the pace and the proliferation of the data that we're we're seeing, whether it's you know wearable devices, but I mean there's just so many different ways that are just generating a lot more data today than we've ever had access to. Can we talk a little bit about the knowledge layer and how do you think about it in terms of what your your platform is is doing, but then more importantly, how do you work with an organization to help them to understand how to manage it and who owns which components of it? So I think there's the the local understanding that you you've mentioned, but then there's this larger thing that you actually have to look at as your your organization is spreading and establishing a much larger knowledge base across your clients.

Declan Kelly: Yeah, 100%. So I think I think that's the difficult side of this two-sided problem that we have tests and solves. One is managing it, the other is accessing it. The clinicians need to be able to access it within a few seconds. If it's managed well and made connected to the tools that their clinicians are already using, they will be able to access it in a few seconds. So that's that's still a hard problem to solve, but not as hard as the managing, because the managing of it is so multi-stakeholder and it's so multi-factorial. It's not just a it's not a clinical problem, it's not just an operational problem, it's a legal and governance problem as well. And the the other issue with it is about roughly about 10% of this knowledge is actually codified and written down, and the rest of it is is in people's heads or doesn't exist. So the ability to capture that knowledge, that what we call that tacit knowledge, the ability to capture that tacit knowledge at scale is another thing that the platform is is is meant to do as well. So it captures the knowledge that exists in clinicians' heads about how this actually works in this health system. So it's only through a combination of both indexing the codified policies and making sure that people have owned them across the business units, and then capturing the tacit knowledge as well, and then making sure that that is seamlessly accessible by all of the point-of-care systems, so within the ambient listener, within the EHR, within your instance of OpenAI or Anthropic or whatever model that you're using itself, et cetera, to connect all of those points. And that that that's really the difficulty. There's a there's a diagram on our on our site that basically shows all of the potential areas where knowledge can come from a health system. And then you know, Eolas is in the middle. We capture and index all of that and make it AI ready, but then there's all of the potential areas that it should be surfaced to, which is you know, within the EHR, within you know UpToDate or any other platform, etc. So that's that's really the two-set of problem that we're that we help health systems solve.

Chris Hutchins: Right. I want to kind of get into one that's being debated a lot here in the in the US, but really in the regulatory space, because it's a whole it means something completely different for an organization like yours that's actually trying to support a global community here with what you're doing. How do you think about the really hopefully kind of influencing the public policy discussions in the places where you have the opportunity to do so, whether you know whether it's in the debates that go on here, whether it's federal or state. At the end of the day, it's gonna be one or the other, maybe a combination of both. But what are the basic things that that you try to make sure that people are paying attention to and that you would be encouraging people to dig into so that they're thinking about it in a way that we're dealing with a some level of a standard as a base level as opposed to the all the the things that they're gonna be fighting amongst themselves over who gets who gets to decide. That's one thing. But being on the receiving end of it, being able to implement and manage to it and actually continue to do the right things for the population that you're trying to serve. Talk a little bit about the challenge for that that you have there and how your how your organization thinks about that and how do you work with your clients?

Declan Kelly: Yeah, so I think I think that this is it's a very it's the same conversation with with different neurons in in every geography because of regulation, because of appetite and adoption of AI and expectation. To be honest, being able to move fast and innovate and get technologies adopted is actually much easier in the US than it is in in other geographies which are more heavy on the regulation side of things, which is the first thing. Zooming out in general, I would I would we I initially take the stance of this is an extremely powerful technology with an enormous amount of potential to do an enormous amount of good in the world. Healthcare is fundamentally broken in lots of aspects of it, and we need deflationary economics in healthcare. Healthcare needs to be it needs to be better, faster, and cheaper essentially, and more people need to be able to access it. And technology has shown in every other industry that it has deflationary effects. Unfortunately, to date, that is not true in healthcare, especially in US healthcare. More tech, more scans, more blood tests, more drugs has actually inflated the cost of healthcare without really significantly improving outcomes at that level. And that's the lens through which all of this needs to be seen. We have an enormous potential and enormous opportunity here. So, you know, whether it's regulation or whether it's adoption of technology needs to be taken through the lens of that. Now, in saying that there has to be clinical safety in mind, you know, this these are going to be parents, brothers, sisters, brains, etc., they're going to be treated with this technology, and we need to make sure that we have we are we're responsible to make sure that it does net good in the world. And approaching that in a responsible way is really, really important. We have a very, very high threshold for clinical safety. And we have a team internally, and every single question that ever gets asked gets reviewed by a separate answer engine to make sure that it was clinically safe and flags it us for any for any issues. And building that culture inside the company is really, really important. That same culture then needs to exist within health systems that are adopting this technology. And if we're able to work on that both together, both you know on the vendor side and the health system side, that the that that means the clinical safety bit is is looked after so that the upside can then be can then be gleaned. So that's kind of generally the the the view that we take it at a very high level.

Chris Hutchins: Yeah, no, it was something that you you've written a fair amount about, and you know, you just kind of leaned to you leaned into it just a little bit there. But really it's it's the transparency and how do you enable a clinician to really understand the source that the recommendations are coming from? So like I mentioned the you know, the open AI or your ChatGPT example of patients putting prompts into the thing, they're gonna walk into the doctor's office, probably have the outputs with some recommendations. Showing how how important is it from your perspective to be able to show the clinicians what what is behind the conversation or the recommendation, excuse me, or the AI answer that's being delivered. How important is that and how do you think about it in terms of how you're how you're pushing the envelope to do that? Because I I don't know that that's an easy thing to solve, but I think it's something we actually have to do a really good job of it, or we're gonna have a really horrible adoption problem at some point.

Declan Kelly: Yeah. Yeah, I think I think it's absolutely vital in the absence of heavy regulation. If we compare this to a drug, for example, a doctor will confidently take a drug off the shelf that there's an enormous amount of evidence for, and that's being cleared by the FDA because they know that that's gone through 15 years of clinical trials and rigorous approval from independent approval. These AI models, you can take a model off the shelf, you can fine-tune it, you can get a knowledge base, and you can give it to a doctor, and you can build that all in the space of a month, and then everyone can use it. And if you look at OpenEvidence, for example, they went from zero to eighty percent of all physicians across the US using their platform, and that was in the space of a couple of years. And really, there is no regulatory body that has that has said, hey, you have to do 10 years of trials and you have to prove all the clinical safety, etc. Etc. Etc. Therefore, trust and transparency are absolutely paramount there. The ability for a clinician to independently verify where the answer has come from, to be in the loop to actually then make that decision themselves is is is is really really vital. And every each country is moving in different directions here. You know, there's some countries that are almost putting the brakes on this so that so that they can say actually this these answer engines are going to be class one or class two devices. Ambient listeners have been labelled as as class one and in some cases class two a devices in in Europe, and then they need to go through the necessary processes and and together the evidence for that. Now that's not the case in in the US. So you kind of you're seeing these strategies at a country level diverge in the absence of any firm regulatory framework that we can point to for trust. The transparency is is is is is is vital really because ultimately that's what the FDA or or the MHRA or the EU act as. They act as a surrogate marker for trust so that you can say, hey, this has been through rigorous trials, I trust this because of that. And even at that, doctors still go through the process to make sure that they're giving the right drug to the right person for the right indication, etc. So there is an absence of that in AI at the moment. And I my honest feeling is I don't think anyone in the world knows what the right approach is. I don't think they get the balance between this innovation and the regulation. And there's a lot of to in and fro in and back and forth and a lot of geographies at the minute of what is the right approach here.

Chris Hutchins: Yeah, of course we're not we're not gonna land on something simple in healthcare. If there are, they're they're probably not they're not the things that we want most for them to be. But I love that you know we have people like you that are thinking about this stuff and pushing the envelope, and we're gonna we're we're gonna see some pretty cool breakthroughs. We're already seeing some, but I I'm excited about what we're what we're gonna be seeing in the even in the next 12 to 24 months because just the pace is just insane how far things are advancing. And I've never seen anything like it in my lifetime for sure.

Declan Kelly: The the phrase of you know, this is the worst it's ever gonna be. I think it's a very a very good phrase, you know.

Chris Hutchins: Yeah, I love that. So what are you kind of take a look at things from an organizational standpoint? As you're talking to organizations that are they know that they have to start making some moves in this space, they're not quite sure which things they have to do first and all that. So as you're having your early conversations with the health system, what are some of the things that you you tell them that they really should be thinking about? They need to get into place so that they're prepared to actually work through an implementation. And and there's a lot of governance and things like that, that's probably mechanics from an operational standpoint that they need to be thinking about. That they may have some bad experiences or no experiences, or they may have some things they gotta tear down or start over altogether. But talk a little bit about that for for listeners who are thinking about what to do to get ready to implement something like what you've built.

Declan Kelly: Yeah. Yeah, we we we we get asked this quite a bit actually, and I I always give the answer of be problem focused, as in what is the problem in your system that you're actually trying to solve, and then try to find a solution. I I actually think technology shifts are the caveat to that answer. And what what do I mean by that? So if if if such a powerful technology comes along, it completely shifts how we do everything. And you probably have to look about that technology and say, we should probably implement this because everyone's implementing this, even though it might not be the biggest problem that we had in our system, but this solution is so powerful that it just completely eradicates this as a problem, if that makes sense. So for example, you can be problem-focused and say, right, you know, we want to take our waiting list down from X to Y. How do we do that? What technologies should we implement? What processes should we implement, etc.? That's a strategic priority and a strategic problem for the health system to solve. And being hyper problem focused is really important as opposed to just implementing the solution for the sake of doing it. The flip side of that is that 10 years ago, convolutional neural networks really hit the scene properly and showed that you can detect strokes much, much faster, you can detect fractures much, much faster. That might have been the top number one priority in that hospital, but suddenly there's a solution that eradicates that problem. So actually looking at the solutions that exist and go, wait a minute, we can probably eliminate problem four from our list because this solution is so powerful, it's so cost-effective that we probably should do that. And very rarely does such clear-cut things kind of come as well. AI is without a doubt one of them. And if I would if you're if you're just asking the question of what we should do for AI, every health system that's detecting strokes or doing radiology should should have AI for radiology. Every health system that does consultations, which is every health system in the world, should have ambient listeners. It the the evidence is a lagging indicator, and even at that, the evidence is now becoming overwhelming. In our opinion, every every health system should have an AI-ready knowledge base so that you're able to surface the knowledge into these ambient listeners and into the EHR so that it improves efficiency and improves the quality of decisions, etc. So there's a number of things there, and there's health systems everywhere in the country that have implemented all of these solutions, and so there's a blueprint to actually follow, and you can scale that up or scale that down to as and when you need. So suppose the short answer is be problem-focused unless you know a sensational solution comes along, and then really ask yourself, should we be looking into that as a solution?

Chris Hutchins: Yeah, things move from it's it's crazy. So we we go from these moments in time where we think we're just dragging our feet, we're taking forever to make these transformations in healthcare, and then all of a sudden we're having a scenario like we're in right now where we can't really move fast enough in a lot of ways. It's it's just it's mind-blowing how quickly the the technology has hit a point where it's it's really forcing conversations that we would have dragged our feet on much, much longer in healthcare, at least from a historical standpoint, for my 30 plus years being around it in the US. Is we're kind of landing the plane a little bit. The what's exciting to hear from that you've been talking about is the expansion piece of it and what Eolas is doing, the the potential for some really massive impact on a global scale, I think is is pretty exciting. So I I love that part of what you're about. Well, you got something that's really as impressive as this is. You do you want to get it to as many people as you can. If it's gonna save lives, it's gonna actually help to hopefully increase the satisfaction of clinicians as well and maybe reduce some of the the burnout concerns. But as you're expanding into the from the you know from the NHS into more US health systems, what's one of the one are two of the lessons from scaling across the UK that you're most worried that might not translate directly?

Declan Kelly: The US in in general is at least five years ahead, if not ten, in some places, with regards to technology adoption. Like for example, all the hospitals across Northern Ireland only moved from paper to their electronic record there over the last three or four years. And there's a lot of scattered systems across the NHS that are very early in their adoption of an electronic health record. And because of that, there hasn't been then lots of systems kind of built on top of that. Therefore, just one of the very pragmatic fundamental differences is where and how clinicians access knowledge is very different in the US than it is in the UK. In the UK, they're used to going to more standalone solutions because the EHRs aren't fully embedded everywhere. Whereas integration into the workflow is way more important in the US. As in these, this needs to be surfaced without them going to the Eolas app. It needs to be surfaced in the EHR, it needs to be surfaced in the ambient listener. That's true in the UK, and the UK is moving towards that, but it's it's it's way, way, way more true in the US because it's more it has adopted technology much faster, basically.

Chris Hutchins: Fascinating. I I guess I don't know how how we get to these moments in time, but it just feels feels like we're we're at this inflection point in history, and it's just fascinating to me that we get to be part of this. And I'm I'm seeing and hearing things in real time that I never would have imagined even a decade ago. What what you're building is is is really impressive, and and I congratulate you on the success and excited to follow you, continue to follow what you're doing with the company and hopefully save a lot more lives, inspire a lot more people to go into medicine, maybe burn out a few less folks along the way. I'll share a great comment I heard from from a physician friend of mine when I first launched the show. He was hit one of the pieces of advice he gave is like if you want to be sure sure that your AI implementation fails, don't talk to the nurses. You'll fail. I thought it was great. But you know, one of the litmus tests I told a couple of people, like, how do we know if we've really missed the mark? Well, just take a walk on the in on any one of the units and count the sticky notes at the nurses' station. If you can count them quick and easy, you might be okay. But you're probably not going to find too many of them because we've done a lot to disrupt some of the things that need to be easier for them. But I again I I just love the the focus on this. And from my standpoint, my entire career has been about trying to figure out the ways that I can support people who took the Hippocratic Oath. Try to make sure that we're doing things that make it easier for you to do what you train for, and that's just to take care of people, help them get better, teach them how to stay healthy, all those things. And you've been doing God's work, and I love that you've just taken this, taking this mantle on to bring this kind of capabilities to really make an impact on the industry. So thank you so much for that. As we wrap, maybe tell people a little bit about you know where where can they find you? I'm sure you have websites and things like that. I'm not I'm happy to share that on the show notes, but just you give folks a quick idea of how best to reach you.

Declan Kelly: Yeah, of course, yeah. So on LinkedIn is I'm more than happy to have a conversation. Reach out to me, and my messages is more than happy to jump on a call and chat. And really happy to understand you know what problems healthcare leaders are facing in their systems, and if we can help and just really just want to understand more. So LinkedIn's a pretty decent place. Email is very good as well. So just on the site, so hello@eolasmedical.com. Feel free to reach out via email too.

Chris Hutchins: So thank thank you so much for for coming on the show and you for just sharing your enthusiasm. And I I certainly feel a lot more confident and encouraged from this conversation, just understanding how clear thinking you are and and what you've got your your teams motivated to do is quite impressive. So again, thank you for being on the show. And I I can't wait to be following in and seeing what you guys do next. I I know that you're gonna be on the shores here and doing some good stuff in the U.S. So I hope we get to bump into you at some point as well. Thank you so much for my listeners. Thanks for listening to this this episode. And if you enjoyed what you heard, you want to hear more, click click the like, subscribe, drop us a review. And we'll see you next time on The Signal Room. Bye.