AI in Life Sciences: Balancing Risk and Innovation

0:00The use cases for generative AI are kind of like a mirror. If you know yourself really well, you have that self-awareness as an organization, then you can really help understand what can be solved with regards to generative AI or really any innovation. Welcome to the Talking AI Podcast, where we talk AI with both experts in the field and early adopters. I'm your host, Matt Page, and we're here to demystify AI for you so you can get some value from it. Let's talk some AI. We talk to experts in AI and early adopters on the Talking AI podcast, and today's guest fits both those criteria. So it's going to be fun.

0:37You're in for a treat today. But today we're joined by Matt Lewis, founder and global chief AI officer at Inizio Medical. And Inizio Medical's focus is in the life sciences industry, including pharmaceutical, biotech, medical device, neurotech, visual therapeutics, all kinds of really interesting stuff. And they're huge. They have 15 ,000 people, 41 offices worldwide, and they're the commercialization partner to a lot of the big names out there in the space. But welcome to the show, Matt. Thanks, Matt. I'm so happy to be here. Appreciate it. Yeah, excited to dig into this and how Gen.AI is impacting the life sciences space.

1:15And like with any industry right now, the promise is Gen.AI is going to completely revolutionize it in the life sciences space, and healthcare is no exception. However, in this industry, the risk of being wrong is elevated in a major way because there's real consequences. And that's exactly what we're going to get into today in the episode, along with the challenges, the breakthroughs, and really all the interesting stuff Matt and his team have learned thus far in their adoption of AI in the life sciences space. But Matt, give us a bit of context before we jump into it of your team's makeup and background and kind of your role as chief AI officer to set us up.

1:55Sure, sure. Happy to do it. Again, thanks to the audience. Thanks for taking time with us today. It's a real passion of mine to try to use augmented intelligence to really speed time to decision. Life sciences is one of those things that, you know, everyone touches one way or another, and if we can find a way to improve people's health through AI, you know, it's so much better for all of us, for our friends, our family, et cetera, et cetera. So thanks for giving me the time and the floor for a little bit. Just a point of clarification, though, I did found the AI, Organized Intelligence Practice here at Onesio Medical.

2:24And before that, the Medical Analytics and Innovation Group, which is our kind of medical affairs, advanced analytics practice across the group. I am not a founder of Onesio Medical itself. I wish I was. It would be a great benefit if I was a primary equity holder of the company, but so much not the case. I'm just an employee like many folks. Good clarification there, but like initiating these AI practices and spaces within the company, That's not a small undertaking, not a big deal. So we're not going to shortchange you there. Sure, sure. Yeah, I mean, so maybe how I got here, if that's helpful.

2:59Yeah, maybe a little bit about how you got here. I think, too, like your team's makeup. Like, what is your team composed of in this, you know, role as chief AI officer? Who are you working with in that regard? Yeah, yeah, sure. So the chief AI officer role is a new kind of new-ish position, if you will. I've been enrolled for just about a year and a half now or so. It was appointed late spring 23. So I have a chance to get my feet wet and kind of figure out what it is the organization wants me to really be doing. But there is an internal expectation to help transform the Inizio ecosystem across the way we work and kind of how we work smarter, as it were.

3:40And then there's an external disposition as well to help our clients and our partners and patients and clinicians and payers and regulators do smarter things. So internally, that's all about thinking about kind of juxtaposing generative AI on top of our operations so that we can be more efficient and productive and make smarter decisions. but externally there's also a consideration of kind of what does good look like and how do we do that in responsible and ethical and compliant ways so that we can speed time to content and so that people that are really challenged with different health decisions can optimize their situation and kind of really improve outcomes.

4:18So I have both kind of responsibilities and depending upon kind of what time of day it is and what day of the week it is and which country I'm in, it's a different work stream typically that I'm engaged with. So for example, it could be that you know, part of our responsibility is working with, you know, our learning and development colleagues. I have a colleague who leads the learning and development sleeve within Inizio Medical. She's based in the UK. And so, you know, we have a curriculum across our 2 ,500 plus staff internally to upskill or kind of augment the capabilities and competencies of our staff.

4:51We're mostly PhD, MD, PharmD staff, if you will, very technical, deeply, you know, intelligent folks, but learning from a digital skills perspective how to use AI to write different types of content or work with experts in ways that maybe haven't always been the focus, if you will. So that's one aspect of the work. Another aspect that I work with our chief transformation officer was Rich Lawrence to think about standing up pilots and experiments across our organization where we can solve for some of the key strategic challenges across the organization and thinking how we kind of optimize for that, those issues that we can solve internally and figure out what needs to scale, what we can deprecate, how we can advance, you know, really that learning internally before we kind of take on a bigger kind of chunk, if you will.

5:46We have hundreds of pilots that we're kind of exploring at any given time. Quick break in the pod. If you're listening to this podcast, chances are you've been thinking about how to actually use AI inside your business. And that's exactly why we built the AI Opportunity Finder. It's a free tool that helps you uncover high impact, tailored AI use cases based on your business, your goals, your pain points, and your industry. No fluff, no generic use cases, just real ideas that fit your business and the rank by ROI potential. It takes about three minutes to run and it's like having your own personal AI strategist for free.

6:20If you want to try it for free, check out the link in the show notes or go to hatchworks.com backslash AI-opportunity-finder. All right. So I got several rabbit holes. Kind of want to go down now. But there's two things that I was picking up there. And we see this a lot with folks we're talking to and just seeing in the market and all kinds of reports that are out there. But there's this concept of, you know, how do I apply it internally to my operations and how I operate my business? And then how do I integrate it into the services, the products and things I'm offering? There obviously is a kind of a risk factor that comes into play.

6:58And then you also mentioned that you have all these different pilots going on. What does that process look like to identifying what are we actually going to pilot? What does that look like for you all in your organization? How are you prioritizing those? because I think it's super interesting for folks because a lot of people are sitting there at this huge Excel sheet of all these use cases and things. It's like, where do I even start? Yeah, I mean, we do this work internally. We've been doing it for years, years and years and years. And we also do it on behalf of and with our direct clients. So, you know, I was with an organization this morning.

7:35Well, it was their afternoon, but it was my morning. And they want us to do this exact type of implementation across their commercial ecosystem. and they were discussing with me, what should the use cases be? Should they be things that come right out of McKinsey white paper? Should they be things that our leadership is telling us are important? Should they be things, Matt, that you've heard out in the world that you think are resonant? And I said to them the same thing I'll say to you, which is that the use cases that an organization needs to satisfy are really unique to that organization. There isn't like a magic list of use cases that works for everyone.

8:10It's really like the use cases for gender of AI are kind of like a mirror. You know, like if you know yourself really well, you're having that self-awareness as an organization, then you can really help understand what can be solved with regards to generative AI or really any innovation. Because like for an organization, for example, that has just come out of a major restructure and, you know, has shifted the boxes on the org chart and there are costs to be kind of extracted, if you will, the things that they need to solve for with regards to, you know, their strategy, implementation pilots and the rest are going to be very different than an organization that is in late phase two and clinical trials for a new drug has not gone through a restructure and is looking at kind of speeding time to market.

8:54Very similar kind of entities on the stock market, if you will, but very different internally. One is focused on, you know, efficiencies and gains. And how do you think about, you know, kind of optimizing for headcount and, you know, helping with productivity? Others much more focused on content and omnichannel and those types of considerations. So, you know, I can't, you know, kind of as a, an external kind of officer come in and say, these are your priorities. Only the leadership knows that, but, you know, when you compliment that with an understanding of what AI can and can't do together, it's magic.

9:26Is there any type of, you know, matrix or assessment you take folks through, like, obviously they're the expert, you kind of want to extract it out of them. Is there anything you're looking at in terms of whether it's the value, the risk, are there other attributes and parameters you're looking at at the use case level as you're trying to figure out how we prioritize one use case over another? Yeah. Yeah. I mean, there definitely are a number of kind of frameworks and, you know, kind of different models or kind of mental models, ways of looking at assessing use cases and thinking about like what's really kind of the paramount Ford organization, all the big firms, the BCG, McKinsey, all have one version or another of these types of works and they all use different words, but they all essentially say the same thing using different words.

10:21I believe the McKinsey model is one where they suggest that firms that have a kind of an opportunistic kind of consideration about what needs to be solved, that they kind of approach that from what they call like a taker mindset. Like, you know, they need to kind of take the available like licenses or software considerations right off the shelf. Like, you know, Microsoft Copilot is a great example. It's right there for the taking. Just get a license. And now all of a sudden people can use Gen.ai directly without having to learn how to code. They don't use any models or algorithms. They just use it directly.

10:52And all of a sudden AI is working in their organization. That's like the easiest way to kind of parse Gen.ai. The next stage is what they call shaping or shapers. They're working with firms that are out of the ecosystem like ours and others to collaborate on platforms or models they built. And those platforms have some maybe configurations, if you will. They're not highly customizable, but they can be adapted a bit to use. And that's a more advanced consideration than just kind of buying a license off the street. And then the biggest thing that people are doing, not a lot of people, but some people are doing like, you know, the Bloomberg's of the world, you can work in Chase's of the world, Verizon, other companies like that maybe are like, you know, what they call a maker.

11:33Like, so, you know, they're building their own models and they're from scratch, like designing things that are just for their own environment. I think BCG calls the same types of things that McKinsey calls taker, shaper, maker. They call them deploy, shape, and reshape, and invent. But it's the same consideration that if there are things you can do cheaply, get experience. Understand what works for you and what doesn't. If there are things that you can partner on or buy that add value in the short term, you should definitely consider doing that because there's no substitute for experience. But some organizations are going to spend a lot of money to build something that's unique to them that they really own as their own IP.

12:12It's not common, but it does happen. And there are some groups that think that that's important. That's a good breakdown. So back to your point earlier, looking in the mirror, you kind of have to know where you are in the marketplace. What type of organization are you and where you're trying to play in essence there? And you've had a couple of years or so to play around with Gen.AI tools as they've been coming out and proof of concepting and whatnot. But in this life sciences space and category that you live in, what are some of the interesting use cases or problems you've seen arise that Gen.

12:51AI is uniquely suited to solve any that come top of mind, whether they're internally faced facing or if they're kind of external solutions that are emerging? Sure. Yeah. I mean, there are a bunch. I'll share two that I think are probably resonant for kind of our world, if you will, because, you know, there are a lot of challenges in the space that we're in to get, you know, relevant, contextual, deep subject matter expert content to the audiences that matter, if you will. The two biggest ones I think that people talk about a lot are what's called like a patient or provider or kind of lay person summary, essentially, which is like a quick kind of overview of a clinical trial that is provided to a lay audience.

13:37In the European Union, this is a regulatory requirement. When you finish a study, you have to publish this type of thing. In the U.S., it's not a requirement, but some companies still do it. So if you're a patient, you're on a particular drug, it's kind of the only thing you have that says this is what the drug has been studied to do and what the objectives were, the endpoints, what the results were, and what you should kind of consider for it. But because there are so many things that go on to get out of the trial and into the market, it's often one of the last things that gets done. So to get from the backlog of the study completion into the actual environment, it takes a lot of in-depth technical writing, a lot of review, a lot of editing, a lot of annotation, a lot of approval.

14:18So, GenRV can kind of speed time to completion by helping to summarize, helping to edit, helping to help kind of provide an early draft potentially for a subject matter team, people to understand what patients' lived experience is like, to stand up a really kind of working version that instead of taking 30 days to complete 30 business days or essentially like two and a half, sorry, one and a half months, you can do the whole thing in less than a week. so you can get potentially a much shorter time to delivery and the quality doesn't suffer so if you're in the uk and you need like maybe a sixth grade reading level or such and you can still produce that in a much shorter time and you know have a quality that people can respond to and hopefully make decisions so that that's a good thing i when my apologies i was i was going to say that when when people first started doing this though i think they said okay why don't i just use like you know chat gpt and i can just you know go and build this direct and i think that's a it's a great question and it comes up all the time i encourage people to do that go into chat gpt or you know gemini or anthropic or you know whatever people use and try to solve your problems locally because when you get to a point with gen ai on your own and you get that like frustration where it doesn't do exactly what you want it to then you kind of realize kind of what the limits are of the technology.

15:42There are things that certainly can be done off the shelf, but most of the really kind of purpose-built applications that people really need solved in the enterprise are not commercially available through consumer applications like JetGPT or Gemini. So you really do need more of a specific implementation, like the ones that are being built by firms like ours. So it's helpful to build that experience. You can kind of understand what the limits are, how to prompt, what a temperature is, and all the rest. But you get to a point you're like, hmm, this isn't exactly what I needed. And to get what I can actually share with patients or with the regulatory authorities, I need a partner, if you will.

16:17So two things for the audience there that I thought were really insightful is a lot of times folks feel like there's this barrier where they're not equipped to go and start playing around with these tools. And like for me and you, Matt, we're living it every day. So we're kind of used to it. And we have this artificial view of, oh, everybody's using this stuff, but they're not. So I think that's a great point. Just start using the stuff off the shelf or just everyday use cases, tasks, workflows, things like that. Because to your point, then you figure out where's your breaking point. And then, okay, now you can go deeper and figure that out.

16:53The other thing that was interesting, there's this component of personalization that's really unique, I think, with this type of technology. Because you mentioned something about different grade reading levels and being able to tailor the response for the individual. I think that becomes a really unique component. When you're thinking about Gen.AI solutions, it's so much easier now to actually tailor something to a user than what it used to be in the past. You don't need this huge kind of if-then crazy tree of scenarios and things that may happen. You're kind of just plugging in the large language model there in a sense.

17:35So I do want to get into more use cases, but you hit on something like a type of solution when you get to this breaking point. And we've discussed a bit in the past RAG, Retrieval Augmented Generation. I think y 'all are playing around with some of this. I'd love for you to go a bit deeper into how you're using that and if there's other examples of using the technology outside of RAG that you're experimenting with as well. Yeah, I mean, I think anyone that's building in this space right now is using RAG. I mean, it's kind of like the de facto kind of band-aid, really. It's not a solution. It's like a temporary fix until the space evolves to a point where people can really build their own custom models with their own data in a secure and compliant fashion.

18:22And it's a little bit more kind of approachable than it is at present. But it's a lot better than the way things were 18 months ago, for sure. And I'm sure I've seen I've seen a lot of the early readouts from like Stanford and MIT and Cornell and other places where they're suggesting what the next version of that is going to look like, which is a lot more, you know, holistic and a lot more approachable. And what 2025 is going to look like will be really interesting and exciting. I mean, in addition to RAG, the other big problem is that a lot of the platforms people are building are not harmonized to a given platform or model.

19:00So one platform might be built with LAM, another platform might be built with OpenAI, another platform might be built with Gemini. So if you learn how to prompt as a user in one model, you only learn how to prompt in that model. It's not really transferable to another platform, which in large enterprise gets really confusing because if you're like a real person, a human, a real actual human being, and you have a real job and your job is like medical writer or you're a strategist or you're a client services person, and AI is like 4 % of your job and you just got trained to do prompt engineering in GPT-4O, and now you get thrown a Gemini built platform that is completely different, doesn't respond the same way.

19:39You just don't have the time to upskill in that platform on top of all the other stuff that you have in your day job. So it's like this lack of harmonization and prompting between the platforms is like a technical problem that the AI engineers know they need to solve. But in the real world, it creates real challenges for actual people trying to do work. I think eventually this discrepancy will go away and that prompting will be less relevant than it is now. The same thing with RAG will be less relevant than it is now. and what will replace it will make it a lot easier to build. You'll see things more like what Anthropic is doing and what the GPTs are within OpenAI and the rest where you won't need as much technical know-how and be more subject matter expertise that wins the day.

20:20But for now, it is the cornerstone of most of what was being built in couplet with deep subject matter expertise in the spaces that you're building. You can't really build anything if you don't know the space well. just like you can't really get a response or a decision out of a model with prompting if you don't know humans well, if you don't know how people live and what frustrations and friction they have in their lives, you can't just show up at ChatGPT, as you know, and ask it for the answer. It doesn't even know what questions to ask if you don't tell it. So you have to have a sense of the world, so to speak, or you don't get anything valuable back.

20:58Yeah. It's interesting you called it a band-aid. I'm trying to figure out a time in some previous era where there was something similar that happened, maybe in the dot-com explosion or things like that. Because it's an interesting perspective because it is something that kind of gets you there. But solutions will probably evolve over time where we may look back on RAG and think, oh, you know, that simplified approach we were using back in the day. But for the audience, just to hit on what RAG is. So and I'll speak to it and you correct me where I'm wrong, but effectively, you're connecting an LLM to your own documents and data databases, resources, all kinds of data sources.

21:41You're using an embedding model to vectorize that data into a database. So the LLM can effectively query your own data versus the entirety of the Internet. and you get some general response. Because what large language models are not trained on is your own proprietary data. So it kind of gives you that ability to hook an LLM into your data and you can tailor it in a much deeper way. You mentioned temperature and some other things like that. Yeah, and it's one of the ways, RAG is one of the ways in which you can lower the likelihood of responses coming that are not helpful for the intended outcome.

22:23Yeah. You know, I know the words I say, I am not saying that you get responses that are inaccurate or false or wise, as people say. And I'm not a believer in that approach. I think, you know, hallucinations are less a flaw than they are a feature of generative AI. Like the hallucination is intended by, you know, the builders originally to be the product of the platforms. Many folks with whom I collaborate in the educational world and higher education and universities and other settings are using hallucinations productively to help learners, students understand the full diversity of options that arrange from a given scenario and then help their students think about if I were to encounter this situation, what might potentially happen?

23:09What will likely happen when a human does this and what else could potentially happen? And it's a way of kind of encouraging systems thinking. It might not be what most people do, but it's very helpful to encourage people to kind of think outside the box. But, however, if you're talking to a patient about what actually happened in a clinical study, you need to lower the temperature to almost zero and get to a point where the data that comes out is really consistent with what actually happened in the study. Get to a great level that they can understand. And, you know, perhaps more importantly, not allow the conversation to kind of if there is a conversation, you know, they have a chance to ask questions of their data, not allow it to go kind of off topic.

23:48So, you know, it's it's less a question of kind of accuracy, like is the thing that I'm being asked truth or not, which is important. But also, I don't want to ask other questions that are not germane to this interaction, because typically those conversations are appropriate for another interaction, maybe with my doctor or my mother or my sister or my caregiver, but not for the model to encounter. or it might be that the training that's done for the model is really strong on how the trial was designed, the endpoints, the results, the interpretation, and the outcomes, but less strong on other things.

24:26Like for example, the pathophysiology of the disease, like why the disease originates in the first place, the incidence of prevalence. So if the patient pushes on those topics, they're going to get a very superficial response. Like they'll get like, for example, say it's in, I know mental illness very well, say it's in like depression, they'll get an answer like one every five people has suppression they start asking more questions on that there isn't data in the training set to be able to respond to that query and they'll find that their answers are left wanting because there just isn't the content to respond to their their inquiry so you don't really want to go too far down that path because they're really just there to get a summary of the trial so you have to really construct the set in such a way that what the conversation is really aligns to their need and to their your expectation so it's both accurate and helpful assistive if you will you start getting into this new approach to i believe user experience uh interacting with an llm yeah a lot of the principles are going to stay the same but there's new considerations to the point you're just making and for the audience i got to go back to the point you just made, which was super relevant.

25:36And I think this is where a lot of people bang their head up against the wall. But you mentioned hallucinations being a feature, not a bug. That's critical for people to wrap their minds around. That is the benefit of LLMs, that it can do this kind of thoughtful approach to thinking through things. And yes, it may not always be quote unquote accurate like we're talking about, but that's through finding the right use case for the right technology is critical in essence. But I just wanted to call back to that because that's, that's a great point for the audience. And then you mentioned one other thing too.

26:14I want you to. Matt, if I may just on the feature and off law thing, I mean, I think other, I've seen other AI experts comment on this as well. I mean, I think it's how you use it and when you use it. So it is that hallucinations, you know, as they're called, you can be helpful when, even when what they say doesn't come to pass, It's essentially Jenny I is a prediction machine. And, you know, sometimes the predictions will come true and we call that accurate. Sometimes it doesn't come true. We call it false. But it doesn't mean that what it said is wrong. It's just it's a prediction, you know, wasn't realized.

26:45But it's kind of like if you're planning for war, right, if you're going to war, there's a time for planning and there's a time for doing. But when you're planning to go to battle, planning and strategy is paramount. paramount but when you're actually the troops are on the field it's no longer a great time to be planning how to mount the you know the infantry because they're already marching down the field so that's not a great time to be like starting to plan but that was there was a time for that and that time has passed but now is the time for implementing when you're implementing you've got to turn the temperature down to almost zero so you get implementation high and planning low but when you're planning you want a range of ideas you want to brainstorm you want to think about all the crazy things that people tend not to think about.