2895: Elsevier - AI in Healthcare: The Role of ClinicalKey AI

[00:00:00] Is the future of healthcare already here? Today we're going to dive into the world of clinical

[00:00:07] key AI. My guest today is Rhett Olden, CTO for Health Markets at Ellspear. And the talk promises

[00:00:16] to revolutionise the speed and the accuracy in which clinicians can access critical medical

[00:00:22] information. And we'll discuss the integration of generative AI with Ellspear's vast medical content

[00:00:29] and how it's creating new possibilities for clinical decision support. And also hopefully

[00:00:34] answer questions like, how does this technology manage to enhance clinical decision making

[00:00:40] while ensuring reliability? And of course, most importantly, transparency. It's a huge topic right

[00:00:47] now. So let's find out more from Rhett himself. So buckle up and hold on tight as I beam your ears

[00:00:54] all the way to St. Louis, where Rhett is waiting for us today calmly, even though it's 95 degrees

[00:01:00] Fahrenheit there. But enough rambling for me. Let's get Rhett onto the podcast now.

[00:01:08] So a massive warm welcome to the show. Can you tell everyone listening a little about who you are

[00:01:14] and what you do? My name is Rhett Olden. I am the Chief Technology Officer for

[00:01:20] the Ellspear Division of Health Markets. I've been there a number of years now, but I've spent

[00:01:25] my entire career in health care, leading technology and developing solutions for patients and clinicians,

[00:01:32] which is what I do now. Well, it's a pleasure to have you on the podcast. And one of the things

[00:01:37] I always try and do very early on in the podcast episode is find out a little bit more about the

[00:01:43] story or the origin story. So can you tell me a little bit more about what inspired your work

[00:01:49] there to develop this advanced clinical decision support tool and what makes it stand out from

[00:01:54] existing solutions in the health care industry? Because there's got to be a story there, right?

[00:01:58] Yeah, maybe part of that story is why I'm at Ellspear, right? Because I spent my career

[00:02:06] in pretty big companies like GE Healthcare and some pharma companies developing solutions. And

[00:02:11] one thing we always struggled with was clinical content knowledge, right? You could build great

[00:02:18] solutions, but if you didn't have the backing of really robust clinical knowledge assets,

[00:02:25] clinical pathways, a robust decision-making tool around drug interactions, et cetera,

[00:02:31] it's become the challenge, right? If you really get something that's really impactful.

[00:02:37] So Ellspear is in its heritage a publisher and has immense repositories of clinical knowledge

[00:02:47] from peer-reviewed journals, books, articles, all sorts of things that we could tap into.

[00:02:53] And that's what attracted me to Ellspear because in my career in healthcare technology,

[00:02:59] it's this marriage of clinical content, robust clinical content, and technology that really

[00:03:06] excites me because I think we can solve some really hard problems and decision supports one

[00:03:11] of those. I mean, it's really challenging when you think about the terminologies,

[00:03:16] the complex interactions, and particularly now with the acceleration of medicine,

[00:03:23] it's just increasingly complex for clinicians. So having those assets available is a real game changer.

[00:03:30] And one of the things that put you on my radar was clinical key AI. And we are in a world at

[00:03:35] the moment where everyone's going crazy for AI. So what I always try and do is demystify some of

[00:03:40] that world, talk about it in a language everyone can understand and solve some of those real world

[00:03:44] problems that we're talking about here. So how does clinical key AI specifically enhance day-to-day

[00:03:50] operations and decision-making processes for clinicians? And maybe if you've got a real world

[00:03:57] example or case story just to bring that to life as well.

[00:04:02] Yeah, I think a simple way to think about this is speed the question. So in a typical

[00:04:12] environment, a clinician might be dealing with a patient that has cardiovascular issues, diabetes,

[00:04:20] maybe also may be pregnant. And that confluence of comorbidities creates a really complex

[00:04:29] environment. When you think about drug-drug interactions, contraindications, as well as

[00:04:37] how do you achieve all of that while keeping that upcoming birth healthy? So that's a complex

[00:04:46] problem. And in the studies we've done, we found that for complex questions like that, it can take

[00:04:53] a clinician anywhere from 10 to 15 minutes going through literature, looking for appropriate

[00:05:01] diagnoses and treatment options for a patient like that. With a conversational approach,

[00:05:07] that can be on the order of seconds. And that's because they can, in a conversational way,

[00:05:13] construct the question. What is the appropriate treatment for a patient that has hypertension,

[00:05:21] type 2 diabetes, is currently on insulin and blah, blah, blah. And that kind of question

[00:05:28] can be answered through our system of clinical QAI seamlessly. Now, the challenge, of course,

[00:05:36] is can you do that in a trusted way? Can you have an answer that is going to garner the right

[00:05:44] level of transparency to a clinician? So we've worked hard to use generative AI in a way that

[00:05:52] provides that level of transparency and trust. And the way we've done that is by leveraging

[00:05:58] the content that I alluded to earlier through search technology. The jargon in the business

[00:06:05] would be a retrieval augmented generation architecture. But what that really is, is

[00:06:11] being able to leverage LLMs to search indexed content at its source rather than relying on

[00:06:21] an LLM that would be prone to errors in terms of using its probabilistic or neural net matching.

[00:06:28] So the model that we use kind of maintains the integrity of the source content.

[00:06:35] And we've built that intentionally to maintain the trust and clinical accuracy of the system.

[00:06:42] And one of the things I wanted to highlight is everyone's going crazy for AI at the moment. They

[00:06:46] have been for the last 18 months, we've seen a lot of mainstream adoption. But of course,

[00:06:51] it's nothing new. And you've had over two decades of experience in leveraging AI and machine learning

[00:06:58] at Alpsphere. So how has your approach to technology evolved, particularly in the context

[00:07:03] of creating clinical AI? Because I would imagine that there's been quite a few big technological

[00:07:09] challenges you've faced along the way. Is any of those that you could share and maybe how the

[00:07:13] companies overcame them? Yeah, I mean, it's a good point. I mean, Alsphere, as we said,

[00:07:19] has historically been a publisher. We've also been deeply involved in natural language processing,

[00:07:26] machine learning, data extraction models heavily for decades, as you point out.

[00:07:32] I think what's been new and a bit frustrating with generative AI is that LLMs are a bit mysterious,

[00:07:43] in the sense that they're not well understood theoretically. And they're also not necessarily

[00:07:50] well understood in terms of when and if hallucinations occur. So what we've tried to do

[00:07:56] is really understand the pros and cons and build a system that first and foremost would control and

[00:08:04] manage hallucinations. Because we thought that was one of the biggest challenges in generative AI,

[00:08:10] at least in healthcare. And then also work on assisting to really validate our offering.

[00:08:18] So we have a really robust validation framework that we've built specifically to manage the

[00:08:26] accuracy and precision of the system that we've developed. And we think that's been somewhat

[00:08:32] lacking in the field, really understanding the robustness and the integrity of solutions

[00:08:38] in the wild. As we all know that something like CHAT-GPT will hallucinate, but we don't know

[00:08:46] when or how or why necessarily. We just know that it can be suspect. So what we've tried to do is

[00:08:54] put better guardrails on that within our system so that clinicians can actually have a baseline

[00:09:01] of integrity in the model. And I'm glad you mentioned hallucinations there because

[00:09:08] you're working in an area where information accuracy is paramount in healthcare and there

[00:09:13] is no margin for error there. So how do you at Clinical Key AI ensure that reliability,

[00:09:20] ensure that trustworthiness of AI-generated medical content it provides to clinicians? Because as I

[00:09:25] said, there is no margin for error, people need to trust this for it to be effective.

[00:09:30] That's right. Now a couple of things that I want to be transparent about here is it's never

[00:09:37] zero. There will be hallucinations just by the very nature of the system, but what we try to do

[00:09:44] is provide really a paramount level of transparency for the user. So when you use the system, it

[00:09:54] actually exposes the references that were used to generate the response. And not only that,

[00:10:01] it shows the individual paragraphs that were used to construct the response. So remember,

[00:10:06] I talked about a retrieval architecture. We're not relying on probabilistic matching of next

[00:10:13] word within CHAT-GPT within an LLM. We're relying on summarizing data from peer-reviewed journals

[00:10:21] and plenary publications. So what that means is that for the clinician, they can still refer back

[00:10:29] to the source material very easily. They can check those references, they can validate it if they need

[00:10:36] to and have that level of transparency. And what we found is that our responses are very accurate

[00:10:45] within certain constraints. For example, it isn't necessarily good at calculations. Most LLMs aren't.

[00:10:53] So we inform the user that if you're trying to calculate a complex dosage, this isn't the right

[00:10:59] tool. But if you're trying to really look at differential diagnoses and options for treatment,

[00:11:05] our tool can be very valuable in terms of how it's able to summarize the information.

[00:11:11] And I also wanted to highlight at Elsevier Health, there's a global footprint in content

[00:11:16] distribution and healthcare solutions around the world. So how does Clinical Key AI adapt to those

[00:11:23] diverse healthcare delivery systems and not to mention, of course, the cultural nuances across

[00:11:28] so many different countries and continents? And can it customize information to fit regional healthcare

[00:11:33] practices and patient demographics? Because there's no such thing as a one size fits all here.

[00:11:39] Yeah, yeah. I mean, you're sort of opening up a real interesting conundrum, right? Because I think

[00:11:47] people that aren't in healthcare think that, well, everybody's a patient, we're all human.

[00:11:52] How different is it? It's actually quite variable from country to country in terms of how patients

[00:11:58] are treated, what the protocols are, what drugs are actually legally allowed for sale in that

[00:12:05] country. So there's a lot of variability. And so our content teams manage that. And we have

[00:12:14] specialized content for different regions of the world, different countries, that we can

[00:12:18] inject into the solution to allow that level of customization. Because as I mentioned earlier,

[00:12:25] the system relies on a retrieval structure. So locale can be used as a differentiator

[00:12:31] to pull up appropriate content based on the locale that an individual was in.

[00:12:36] And that customization is super critical, as you point out, because you don't want to be

[00:12:44] making recommendations that aren't quote unquote best practice for a given locale.

[00:12:50] Because A, you lose credibility and also B, that's just not a way medicine is practiced in

[00:12:56] that space. So we're working on that and we have solutions there based on that retrieval system

[00:13:03] and leveraging again the content that we have is localized for that region.

[00:13:09] Now as someone that is an eternal optimist, I've often thought, maybe I'm a bit naive here,

[00:13:15] but I've often thought that if the whole world and all the healthcare systems could begin

[00:13:19] sharing all that information, we could be making so much more progress at solving real-world

[00:13:25] medical issues around the world if we just pulled all our resources and used AI to solve some of

[00:13:29] those problems. But of course, there is a flip side of that. So what ethical considerations

[00:13:35] come into play, particularly concerning things like patient data, their privacy and the potential

[00:13:40] for AI bias, which is another can of worms as well. And how do you navigate these challenges?

[00:13:45] Dr. Michael Sacklett Yeah. I mean,

[00:13:48] these challenges are not trivial to solve. Let's talk first about bias. The advantage that we have

[00:13:58] is that we rely on peer-reviewed journals and articles that have been reviewed by medical

[00:14:04] experts only. We're not pulling information off Reddit or some other untrusted source.

[00:14:15] But that content can still contain biased information, either racial bias, gender bias,

[00:14:21] et cetera. So we have a system to manage that. We've actually adopted a quality management system

[00:14:30] similar to a medical device that actually provide feedback and error handling so that we can

[00:14:36] mitigate those biases if and when they occur. The bias in our system is extremely low or close to

[00:14:44] zero. In fact, we have very little instances where we've seen real bias occurring. But that's

[00:14:51] not to say it doesn't exist. We have to passively monitor that. So that's an area I think where

[00:14:57] we've really leaned in to make sure that we're addressing those societal needs as we go forward.

[00:15:04] Now, the question around global content or how do we better distribute this, I think that's been

[00:15:10] an ongoing mission of Elsevier for the past 140 years. We want to be able to make knowledge as

[00:15:19] transparent and I would say less boundary-less. What I mean by that is

[00:15:29] national boundaries shouldn't matter. We should be able to use scientific information and medical

[00:15:35] information as broadly as possible. So we're also trying to work with plenary healthcare institutions

[00:15:43] in the US and elsewhere to look at how we can move some of their best practices into

[00:15:49] a framework like this so that underrepresented communities or challenged communities can have

[00:15:55] access to the best healthcare in the world, not just the Mayo clinics or the Cleveland clinics

[00:16:02] of the world. Because I think the goal here is around leveling the field, not trying to make

[00:16:10] it even more hierarchical. And I think generative AI can do that. We're providing some of the

[00:16:17] best practice information to any clinician using this, whether it's a small clinic or

[00:16:23] community hospital or a large academic center, which is really exciting.

[00:16:30] It really is. And if the pace of technological change is moving at breakneck speed right now,

[00:16:35] there's also a reality it might never move this slow again. So if we look beyond clinical key AI,

[00:16:41] how do you envision the future of AI and machine learning in healthcare? And are there any other

[00:16:46] areas within healthcare that you might be planning to introduce AI-driven innovations in the future?

[00:16:51] I appreciate you're probably locked down as to what you can share, but is there any other areas

[00:16:54] that excite you around that? Right now we have a tool where clinicians can come to it when they need

[00:17:02] to understand complex questions and gain insights, which is great.

[00:17:09] The next level is to really anticipate that for the clinician. So being embedded in a clinical

[00:17:17] workflow and being able to kind of prompt and almost be a background advisor to a clinician

[00:17:29] as they go through their clinical workflow and stops is the next level. Because we can do that.

[00:17:37] We understand the patient data, we understand the clinical needs, and it's a question of how do we

[00:17:44] work more seamlessly as a partner with the clinician? And so being able to push information

[00:17:51] is the next step. And one area of high interest is oncology.

[00:17:56] You know, oncology is an area that moves incredibly quickly,

[00:18:01] partially because it is so life-threatening. So the velocity of new drugs in that space is very high.

[00:18:07] The velocity of change and protocols and management is very high. It's a perfect area for

[00:18:15] really driving something like generative AI in the clinical space and helping improve patient

[00:18:22] outcomes through that area. And if we do have any healthcare professionals listening, or maybe

[00:18:28] they've just discovered this podcast searching for something around this area, maybe they're keen on

[00:18:33] integrating AI tools like clinical key AI into their practice, but they do have some concerns.

[00:18:40] They might be hesitant due to concerns around technology adoption challenges or impact on that

[00:18:45] clinician-patient relationship and all those separate concerns out there. Is there any

[00:18:51] particular advice to that person listening that you would offer?

[00:18:56] So we've been really clear about people trying it. You know, the best way to use this is to

[00:19:05] try it for a few weeks and see how it operates. And we're more than happy to offer

[00:19:11] retrials to anyone that's interested. And the feedback has been phenomenal. You know, we'll

[00:19:17] have people coming back to us saying, this thing actually improves so much the way I work in my ED

[00:19:25] rotation in the emergency room department, or how it's been so astounding in terms of complex cases.

[00:19:34] So the best way for people to overcome their anxiety and fear is first to try it,

[00:19:41] but then understand and ask critical questions about quality, quality, quality, quality.

[00:19:48] Because I think that within the healthcare space, the most important aspect is can you actually

[00:19:55] demonstrate clinical efficacy? Do you actually have safety protocols in place? And big tech

[00:20:03] hasn't really done that because it will slow them down. And I think this is where

[00:20:09] healthcare companies like Elsevier can make a big difference by really leaning in on the quality

[00:20:17] and clinical efficacy of the solution. So that would be my answer to your clinicians out there.

[00:20:24] Try before you buy and ask some tough questions, you know, around how it actually operates. And

[00:20:31] don't be afraid to ask those questions. It's really critical.

[00:20:34] Matthew Fantastic advice. And for everyone listening,

[00:20:38] myself and indeed you, there is a real pressure on us all to be in a state of continuous learning.

[00:20:44] Now, as someone who's leading the way in this industry, I've got to ask,

[00:20:48] where or how do you self-educate? Anything you could share there?

[00:20:51] Dr. Richard D. Well, you know, it's interesting when we started this journey

[00:20:58] realistically with generative AI maybe 20 months ago, really early, you know, because we saw

[00:21:05] the opportunity, you know, with the emergence of open AI and Thropic and others. The question was,

[00:21:12] how do you use it, you know, for a harder use case? So what we set up was rapid prototyping

[00:21:19] teams to really A, understand the technology better. And then B, we pulled in clinicians

[00:21:26] early into the process during the research and the envisioning, because you need to understand

[00:21:32] your customer, right? And have empathy for what the customer actually needs.

[00:21:38] So we have clinicians on staff. We have two partner hospital systems that work with us

[00:21:45] and we've embedded with them to really build a solution that can meet the needs.

[00:21:51] So I would say that, you know, like a lot of things, be ready to pivot, be ready to change

[00:21:59] and be okay with some failures. Like we made some mistakes, everybody does. But the key point was

[00:22:06] having that target of what does the customer really need? Because that's what we focused on.

[00:22:12] Dr. Richard D.

[00:22:13] Well, I think that's the perfect moment to end on today. But for anyone listening that would

[00:22:18] love to dig a little bit deeper on everything we talked about today, find out more information,

[00:22:23] try before they buy and all that good stuff. Well, do you like to point everyone listening

[00:22:27] if they want to find out more information or contact your team?

[00:22:30] Dr. Richard D.

[00:22:30] Yeah, I would say go to elsevier.com and check out the Clinical Key AI site where you can request

[00:22:40] to download and try the system. And that's the best way to try it. And feel free to ask questions.

[00:22:47] That's where I would start. And also, I'm really excited to see how we can work with you.

[00:22:53] So this is probably one of the most important moments, really since the 90s that I can think of.

[00:23:03] Now I'm showing my age a bit. I've been around in the dot-com era back in early 2000s.

[00:23:08] That was a huge moment of change in technology. I think this thing is as impactful as mobile devices.

[00:23:16] So it has the opportunity to really enhance our lives. It also has the opportunity to screw

[00:23:23] things up. So this is where we really need to be cautious and you need to look for partners that

[00:23:31] really are concerned about doing it the right way. And that's my biggest takeaway here. And I think

[00:23:37] Elsevier is one of those. Dr. Richard D.

[00:23:40] Yeah, I completely agree with you. And a lot of people have likened it to the arrival of smartphones

[00:23:46] and mobiles, etc. But I think it also reminds me a lot of when the app store came the year after.

[00:23:52] When the first app store came out, everything was gimmicky. Developers were just playing.

[00:23:56] It turned your smartphone into a can of beer or a chainsaw or a musical instrument. But then

[00:24:01] after that period, it became right. What are the real world problems that we can solve here? And I

[00:24:06] think with Generative AI, we now have a unique opportunity to create a tool that can act as an

[00:24:11] expert in our world, national and regional health care delivery like what you're doing.

[00:24:15] And what I especially love is you're encouraging people to think bigger of what we can do with

[00:24:20] this technology and solve real world problems. So thank you for sharing that with me today.

[00:24:25] Thanks, Neil. Thanks for having me. Really appreciate it.

[00:24:28] We covered a lot of ground today with around the powerful capabilities of clinical key AI

[00:24:34] from its impact on clinical efficiency to the ways it ensures trustworthy and localized

[00:24:40] medical advice. I think it's clear that this tool is setting a new standard in health care

[00:24:46] technology. But what are your thoughts on the future of AI in health care? The good, the bad,

[00:24:53] the in between, whatever it might be, how do you see advanced decision support tools shaping

[00:24:59] clinical practices? Share your views with me. Join the conversation on how technology continues to

[00:25:06] transform health care. You can do that by emailing me tech blog writer outlook dot com,

[00:25:11] Twitter, LinkedIn, Instagram, just at Neil C. Hughes. But more than anything, just thank you

[00:25:16] for listening as always. Join me again tomorrow. We'll keep exploring the new frontiers of

[00:25:20] technology. But until next time, don't be a stranger.