Large Language Models in Clinical Trials with Eirini Schlosser

This week we sit down with Erini Schlosser, CEO of Dyania Health, and chart the pioneering advancements in large language models within healthcare. This episode promises to reveal how the shift from simple entity recognition to sophisticated reasoning models has revolutionized drug discovery and the curation of electronic medical records, all while unpacking the challenges of data privacy, computational demands, and the hunt for specialized AI talent. Discover the fusion of technology and medicine where the collaborative potential for real-world data utilization in evidence-based studies emerges as a beacon for innovation amidst the complexities of modern clinical trials.

As we explore the parallels of AI training to medical education, you'll be enthralled by stories of AI models undergoing meticulous fine-tuning by our in-house physicians, echoing the rigorous journey of medical residents. Irene lifts the veil on her own path, which led from a lineage steeped in medicine, through the worlds of biochemistry and investment banking, to the groundbreaking integration of NLP into healthcare at Dyania Health. Our conversation shines a light on the future of automating clinical research, where AI not only answers but justifies with evidence, promising an unprecedented era in the management and interpretation of medical data. Tune in for an intimate look at the challenges overcome and the milestones achieved in teaching machines to navigate the intricacies of human health.

Guest:
Eirini Schlosser 

________
Reach out to Sam Parnell and Ivanna Rosendal

Join the conversation on our LinkedIn page

Transcript

Ivanna Rosendal: 0:21

Welcome to Transformation in Trials. This is a podcast exploring all things transformational in clinical trials. Nothing is off limits on the show and we will have guests from the whole spectrum of the clinical trials community, and we're your hosts, ivana and Sam. Welcome to another episode of Transformation in Trials Today. In the studio with me I have Irene Schwasser, who is the founder and CEO of Dynia Health. Hi, irene.

Eirini Schlosser: 0:49

Hello Ivana, Nice to meet you and thank you so much for having me on the podcast today.

Ivanna Rosendal: 0:53

So I'm excited about this and I'm excited about our topic, which today will be large language models in clinical research. So, to start us off, irene, could you tell us more well where are we with large language models and clinical research?

Eirini Schlosser: 1:07

Sure, absolutely so. I think you know natural language processing as a topic within both medicine as well as clinical trials has been an area that has been very much saturated in terms of the phrasing over the last few years. So everyone has heard NLP, nlp as a flag basically being waved. What NLP has meant, both in clinical research as well as globally, has changed quite significantly over the years, and so I think over the past decade, where I like to split it into is basically from the invention of transformers in 2017 and either pre or post that event. So I think healthcare more broadly has been has invested a lot of effort, time and resources into what are called more named entity recognition based type models, and we like to compare these types of models to like a dog hearing its name. So if you're thinking about a patient, note that it's reading, it goes blah, blah, blah, blah, blah, blah, blah, blah, blah, blah, blah, blah, blah, blah, blah, blah, blah prostate cancer. So it's very good at picking up on key words and key entities hence the title named entity recognition but is not able to draw conclusions and inherently reason as a true type of AI would, and so the traditional way of developing NLP in both healthcare, as well as, more specifically, in clinical research, has been to build out, basically, data dictionaries and data mappings for specific meanings, which was an immense amount of work on clinicians, researchers. You have all of these standardization of data models that have been used over the past few decades to try and make sense out of information that is recorded in unstructured, free text and, to be quite honest, where we are today is the healthcare industry is still using those types of models.

Eirini Schlosser: 3:12

There have been an immense amount of investments. There are, you know, there's a little bit of difficulty, let's say, drowning the baby in the bathtub and replacing what is, at a technological level, currently obsolete, and this might be a provocative statement, but with large language models and what that has presented in the healthcare space is an opportunity to really change the way things are done, and the only limitation being the computing power as well as the talent and expertise to train these models and focus on specializing them for various use cases. So I know, you know, we've seen some early successes, for example, like a bridge which generates text. You know we have been, we are very clearly defining and I think the industry is starting to define certain tasks uniquely. So there are some large language models that are very good at answering USMLE multiple choice questions.

Eirini Schlosser: 4:16

There are other language models that are very good at generating physician notes. We've been particularly focused at Diania Health with understanding and drawing conclusions, with the medical accuracy that a human would have while reading an electronic medical record, and so the way you train the models and the data upon which you train the models is still something that is a wild west, let's say, for health care. I don't think that there has been any standardization that we've seen across the board. I think that the main decision makers have not yet had the degree of catching up to where big tech is and AI research is to understand really how to connect the dots with what is otherwise cutting edge research on the clinical side. So that's where I think we are today. That's where we are.

Ivanna Rosendal: 5:05

Well, what are some of the biggest wins we've seen lately in this space? Where have these models been applied and make a difference today?

Eirini Schlosser: 5:15

Yeah, absolutely so. I think in drug discovery that's been quite useful in the fact that models are being deployed to read through research and publications. I think there are a number of use cases on generating notes or generating, let's say, emr, electronic medical records, and so that has been has shown some early successes. I think on the natural language understanding side we're still quite early. Uh, actually, I mean, and that's why we're, we're uh, personally, it's our company, that's why we're focused on that, because we saw that as a very unique opportunity. Um, and I think that in terms of of reading through medical literature, that's really where the, the, the successes next will come out, because I know of a few models that are being focused in on that space, and then I think really the bottleneck to achieving further milestones will be around computing power and the availability of GPU to these larger institutions.

Ivanna Rosendal: 6:20

That makes sense. You recently went to the Scope Summit.

Eirini Schlosser: 6:25

I did, I got back yesterday.

Ivanna Rosendal: 6:28

Well, very recently. I would be curious to learn, based on what you're currently doing and what you saw other companies also doing, where are we going this year?

Eirini Schlosser: 6:43

That's a great question. So I was actually I gave a talk with John Chelico, who's the CMIO of Common Spirit Healthcare System and was formerly the CIO of Northwell Health, and our main focus was around how, within clinical research and specifically real-world data or real-world evidence-based studies, researchers could partner with healthcare systems and what are the roadblocks to that. And I think that's a very unique opportunity that has a lot of red tape involved, and for us, it's been about understanding what are the priorities of the healthcare systems and what are the worries, and so just to highlight a few of these points, you know, the first thing is around data privacy and the IP value of the data. The health care systems do not want the data often flowing out of their health care system computing environment or firewalls. They fear the privacy from a PHI perspective. They also fear IP leakage because they view the EMR data as a value provider within the space, and so I think number one has been an effort. I think the next phase will be an effort to build solutions that fit within their firewall, and it's a shift away from the historical model of data mining and a shift towards AI as a service, or providing a service to solve a specific problem. That's how we've been addressing it and just given input from various advisors who we work with quite closely that are on, you know, the healthcare system side.

Eirini Schlosser: 8:25

On the private provider side, we've been narrowing down on what it how to standardize those processes so that it's not, you know, a unique deployment every time we get to a new healthcare system, and I think that's something that will likely be a number two point on what's next to come, will likely be a number two point on what's next to come, which is a standardization of deployment. So right now, you know, just as an example with regards to cybersecurity, you know we get to one healthcare system and they say, oh, thank God, you guys don't do, you don't require us to do SAML, sso, and then we get to another one that says that's absolutely our protocol requirement, you're not doing anything here without that, and so I think everyone's been having different standard operating procedures as just a you know what their, their previous legacy systems entailed, and so I would expect a standardization of of deployment processes and IT integration to happen over the next, over the next year, of deployment processes and it integration to happen over the next, over the next year. So that's number two.

Eirini Schlosser: 9:28

And then you know, another big trend and focus area is around computing power and the availability of healthcare systems to be able to, to, to run AI systems. So I think you know, I like, I usually like to say, if there's no GPU AI system, so I think you, know I like, I usually like to say if there's no GPU, there's no AI and there are a lot of groups waving AI flags and there's not a lot of GPU available.

Eirini Schlosser: 9:57

So you know we had bought.

Eirini Schlosser: 9:57

We had bought a few clusters of NVIDIA 100s before Elon Musk cleared out the market, but the lag time, even with cloud providers, is months, and for a healthcare system to purchase clusters of GPU, you're talking about hundreds of thousands of dollars worth of machines, as well as then having the talent and resources and housing let's call it data centers to store them, or they're making a shift to cloud, and so I think the last point with this is going to be the next stage of advancements for how healthcare systems interact with data solutions and whether they are on-prem, which they've historically been. Many of them are moving into hybrid environments, now with cloud and in the future, as that becomes likely, almost exclusively cloud. What's the availability of GPU? I don't think that if you have a group like Microsoft that's partnering very closely with OpenAI, there's GPU being deployed for training new models on the big tech side, and so it's going to be an interesting dynamic whether there is going to be availability of computing power for running, inference of actually running and solving the tools.

Ivanna Rosendal: 11:19

I love that it's such a tangible limit to how quickly we can move an AI. The amount of actual computing power. That is mind-boggling to me. What are some? You mentioned that Elon Musk cleaned the market for GPUs. Can you tell me more about that?

Eirini Schlosser: 11:42

Yeah, sure, he spent about $300 million on GPUs and set the supply and demand tipping scale into a little bit of mayhem, has been manufacturing these as fast as they can, and you know that's a whole nother industry. But the problem is global and I think that you know they produce more and then everyone tries to buy them up, but they've been mainly on the big tech side. So most of the people that are directly training these large language models have not been in healthcare and they've not been in clinical research. They've been in big tech, and so I think that's a stark difference. Going back to your question around what other vendors are thinking about, and you know what we saw at the SCOPE conference.

Eirini Schlosser: 12:33

You know there's a new name actually for what the previous NLP was, which was I think it's called precision data mapping, but it's not. You know's a new name actually for what the previous nlp was, which was. I think it's called precision data mapping, uh, but it's not. You know large language models and so I think there's a few qualifying points of understanding whether a group is is has trained in llm in-house, whether they're kind of just repurposing a gpt4 api and what that looks like. It comes out with architectural planning, and so you know, and as well as answering and the ability to answer questions as to how models were trained if they have been trained, I think otherwise. In healthcare, there are a number of institutions that are training models in a more academic setting, either directly with researchers in a medical center, but it's not necessarily being deployed as part of a commercial end-to-end deployment.

Ivanna Rosendal: 13:26

Yeah, I want to kind of look back to the standardizing, the deployment piece of AI. What can be standardized can be standardized.

Eirini Schlosser: 13:35

That's a great question, so well one, I think, the method of interacting with a system and where the data sits and how the data is used. I think that there have been too many efforts in the past to try and just squeeze data out and then sell it and repurpose it in, you know, structured form or de-identified, et cetera. We're looking at a new industry for what data even means. So you know, if I'm looking at a patient population of EMRs and I'm asking how many patients took biologic, that's not a raw data point, that's an insight, and so the insight and the conclusions that are being drawn from that can be taken out of a healthcare system environment without necessarily moving the raw EMR data out. And so I think a standardization around what is classified as data, what is classified as insights, what is classified as IP in these engagements is something that will make any process and educating the right parties and legal much easier across the board. And I think that you know there does need to be an understanding from the healthcare system side that unless they are investing multiple, seven figures into companies that are doing this on a commercial level, you know they're not the real. The value add here is the training of the models and you don't necessarily need to do it on EMR data. In fact, I would argue that in many use cases for large language models, just training it on an EMR data is not really beneficial. You know we view our AI training as like a doctor getting their MD degree. You wouldn't want an MD having read 5 million EMRs, and so you know we can talk about a few ways of how models are trained, but I think standardization of understanding what is X and what is Y is something that will be a first step for healthcare systems and that can come through either. You know, I've heard discussions around kind of like standardization of performance, benchmarking and how AI models can add value, justifications and explainability, auditability, et cetera. So that's one area.

Eirini Schlosser: 16:03

The second is around deployment in actual machines, and I think that's a little bit more difficult to standardize because some groups have their own, you know, high performance computing environment set up. Others are working with cloud, Some have both options and are trying to compare. You know what gets done with others. You know we've set up like VPN tunnels and so the method of how an external vendor can access the information they need and, under, obviously, a business associate agreement, to provide those services to the healthcare system is something that is still a little bit up to each institution themselves. We particularly have worked with what each institution has preferred, and they've often had different preferences as to how that works and where the machines sit, and whether it's cloud or whether it's in their premise, et cetera, and also how the EMR data moves from one part of their institution to another, and the frequency system whether it's caboodle APIs, whether it's HL7, and the ease of use around the format of the EMR data that's being moved from one spot to the sandbox environment of another.

Ivanna Rosendal: 17:27

And that makes sense. I like the analogy of that. What can be standardized is almost like the onboarding of AI, the training and education of it.

Eirini Schlosser: 17:39

That's a great way to summarize it. That's exactly right.

Ivanna Rosendal: 17:43

I also want to look back to AI as a service in general. Is it currently delivered as a service or what is it that will differ once AI becomes a service?

Eirini Schlosser: 17:56

That's a great question. We've been selling it as AI, as a service, and so I think it depends on the use case and how. You know, different models have been deployed, and if it's being provided as a service, meaning it's commercially contracted to provide X, y and Z, I think that you know, other than like the abridges of the world, I think that the majority of AI is being deployed right now in a building process, and so everyone, you know, you keep hearing the term co-development, co-development, co development. Well, do we need to co-develop or are our models already? Is the model for this purpose already performing at above 95% accuracy and it's ready for commercial use end to end? And so I think that's really the question that will start to, you know, come up. Is that, you know, once a model is already trained and it's retrained and it's improved upon and it's performing and benchmarked against human board certified physicians, do we, do we need to continue to have co-development opportunities, or is that, you know? Is that moving out of the academic phase and into commercial setups? And I think that also goes back to the education of of and and kind of development of the knowledge base of decision makers in the space. But for us. You know, I think there's a, there's a specific cutoff.

Eirini Schlosser: 19:19

I would say the majority of the industry is still in the development phase. I think you know chat GPTs very public announcement, kind of in the end of 2022, really kick-started the industry into thinking, oh okay, now we understand what, what AI is, whereas before, you know, you had a lot of groups that were kind of tagging on AI to everything. But when you went and we would actually go through, just so we know who we're working with or who we're working, you know, in a more comparable environment to from a business model perspective, we would go through their, their LinkedIn's and count how many people they had with AI research experience, whether they had publications, how many physicians they had, whether they were annotating or training the AI models and whether they just had, you know, 500 people that had a title of chart abstractor. They've got 500 people somewhere title of chart abstractor. They've got 500 people somewhere. The title of chart abstractor. It's not ai. Yeah, yeah no, absolutely.

Ivanna Rosendal: 20:23

Uh, that's a very nice specific way to put it.

Eirini Schlosser: 20:27

I kind of like to like the, the hamster wheel inside the box. So if it's, if it's a box that's being called ai and you open it up and there's a hamster just running its feet, you know it's not a robot, it's a real, it's a human. The mechanical trick yes, exactly Great.

Ivanna Rosendal: 20:49

I want to go back to what you said earlier about the medical accuracy of your model. I am curious how did you achieve medical accuracy?

Eirini Schlosser: 20:59

Yeah, that's a knowledge base and that you know we spent like probably almost a year cleaning a medical text set for a knowledge base whether it's textbooks, publications, and just building on that knowledge base. So that's why we like to compare it to like a physician getting their MD degree. It's literally they're trained on on that knowledge base. So that's why we like to compare it to like a physician getting their MD degree. It's literally they're trained on a knowledge base. And then the fine tuning for us is actually done with annotated cases by our in-house physician team. So we've got full-time physicians that have cranked out over 25,000 cases, over uh 25,000, uh cases, and that is is quite competitive um, for making the model be very specialized at. And actually we like to compare the fine tuning to like doing its its residency, uh, where it's specialized on performing specific tasks for specific disease and therapeutic areas. Um, so we, you know we had full-time physicians that were training well, basically providing ground truth annotations to reading an EMR asking and generating questions that were relevant for the clinical research areas we were focused on, and then providing a ground truth answer and a justification for that answer. And when that method we trained uh, our our model to be able to explain itself. So it um has to pinpoint back to the part of the medical record that led it to um, uh, derive the conclusion that it that it derived. And so, um, those ground truths uh let's call them annotations from the physician team is what's also used to benchmark. A certain subset of those cases are also used to benchmark our performance against human physicians versus the model. So, yeah, that's how we did that.

Eirini Schlosser: 23:19

Now, that being said, there isn't just, you know, a point blank type of solution to solve for or, let's say, question to solve for.

Eirini Schlosser: 23:29

There's also very we have varying complexities that we will assess the performance against.

Eirini Schlosser: 23:36

So the most complex types of questions are things that have to do with orders of events.

Eirini Schlosser: 23:59

So let's say, you know, we're looking for a patient who is a non-smoker and the you know in in, let's say, today the physician wrote a note and said the patient quit smoking one month before their surgery last year. Then we have to create a chain of thought, literally to go back last year and see when they had the surgery, what surgery happened, and then the date of that and the month prior was when they quit smoking, and so that chain of thought is not something that traditionally large language models are even designed to solve for. So that's an entirely different level of complexity that, you know, my team has been quite focused on solving, but it's very common in medicine. So medicine often has chains of events that have occurred throughout a patient's history and those would otherwise form a longitudinal picture if someone was reading and deducing certain conclusions at each point in time. That's what LLMs today are not otherwise designed, built or capable for solving, are not otherwise designed, built or capable for solving?

Ivanna Rosendal: 25:08

Yeah, that makes sense. Also an interesting journey. I'm curious about your journey into the space, irene. How did you?

Eirini Schlosser: 25:17

end up working with large language models in healthcare? Yeah, that's a great question. So you know, I want to say it starts from birth. I was born into a medical family. My entire family is comprised of physicians. I grew up in the space.

Eirini Schlosser: 25:32

I was digitizing my father's patient records as a high school job, you know, with manila folders and you know, having to pick out which patients were deceased, which patients were inactive, et cetera, pick out which patients were deceased, which patients were inactive, et cetera. So I have, you know, the easy answer is I've been around the space my whole life. I was majoring in biochemistry to also follow into the family business when I decided that I was and not to to, you know, discredit the difficulty, but I was bored with the routine standard of care of after going through this intense process of, you know, going through medicine and residencies and exams and board exams, one after another. Then you know, they, they, I was shadowing and doing clinical rotations with family friends and kept seeing that. You know that most of the days were very similar routine standard of care, which is which is great for the patient. But there wasn't enough of what I was excited about, which was solving a new problem and the and I I'm not a big fan of routine on a daily basis. So at the time, you know, I was a I was a college kid did some applications, fell into investment banking for a couple of years. That was an incredible experience because I was seeing how, at scale, very impacting decisions, high impact decisions, were being made with a limited set of information, and a limited set of information meant that they were, you know, making financial, corporate, wide industry impacting decisions from structured data and not taking into account information that was being recorded in free text in many cases, and so that kind of piqued my interest.

Eirini Schlosser: 27:15

I started towards the end of my time at Morgan Stanley and M&A. I was working on exclusively technology deals and just fell in love with the tech side. At the time I only knew how to code in Visual Basic and doing Excel macros, so I had started moonlighting my first business from 2 am to 4 am and realized I needed to learn Python to manage teams of software engineers. So that was really my start. My first business was also very heavily focused on natural language processing, but in its previous let's call it iteration, which is not really AI. This was before. This is 2014, 2015.

Eirini Schlosser: 27:57

So quite early in the technological development of what NLP is. Development of what NLP is. Long story short. I had some early investors in 2018, 2019, who also happened to own hospitals that were getting acquired by a very large group, and their question to me was their curiosity around the value of EMR data from a clinical research perspective, and so what I had been focused on within the NLP space and as hard as I tried to run from medicine, you know, the kind of perfect storm opportunity presented itself where we and my team and I shifted gears and I probably spent about a year just diving in with stakeholders and down the rabbit hole of clinical research and real world data and where the inefficiencies were that were being just done by human effort and just saw this unique opportunity around how most clinical research coordinators, as well as nursing staff, research nurses and PIs, just spent an immense amount of time doing reviews and reading electronic medical records, and so that process of reading medical records is what we set out to automate, and that's how Diane was born.

Ivanna Rosendal: 29:20

Oh, that's amazing. And is that still your focus? Has your focus shifted? Yeah, nope, that's our focus. So your focus shifted.

Eirini Schlosser: 29:25

Yeah, nope, that's our focus. So we spent the you know we were, we were born end of 2019. We had our, our, our pre-seed funding 20, january 2020, and then suddenly the world changed and that that really was, I would say, um, uh, a blessing in disguise for us, so we could spend a much larger period of time focusing on the technology and came out of the pandemic and where, I would say, the pre-ChatGPT launch, no one really knew. They had trouble in the healthcare space, differentiating an LLM-based approach versus an LP and what that looked like. And then suddenly, you know, chatgpt goes live and you know, within a few months, everyone started to really understand what LLMs meant, and so I would say, yeah, I mean, we've had a fun year this last year, so last year.

Ivanna Rosendal: 30:25

I also love that your early experience with patient data was like the very physical experience of files and papers, like you've seen the data, you have been in the data.

Eirini Schlosser: 30:36

Yeah, my fingers turned yellow from the manila. The data before it had zeros and ones attached to it. Exactly. I love that.

Ivanna Rosendal: 30:47

So it makes data very tangible and I think it's easier to imagine. Well, how else could this look like if you actually understand? Well, what did it look like?

Eirini Schlosser: 31:00

Absolutely a clinician, having to read and understand and, just like, sit there and study medical records in order to make whatever conclusion they were looking to draw for the patient's care, and that mental process is just what. For me, it just seems so obvious that it just needed to be automated. We needed clinicians focusing their care efforts on the patient and on running the clinical research and not on just spending time going through EMRs. So it takes about 30 minutes or more, depending on the complexity of the patient's disease area, to read through a single medical record, and so even chart abstractors. They can get through about 4,000 EMRs per year, and so when you're looking at the entirety of, you know, hundreds of millions of EMRs, that's just an immense waste of time If you can do that on an AI level with a fraction of the time.

Ivanna Rosendal: 32:00

Yeah, absolutely. That has a great potential. Well, which brings us to the question that we always ask, I guess, on this show, and that is if we gave you the transformation trials, magic wands, that can change one thing in the life sciences industry. What would you wish to change?

Eirini Schlosser: 32:20

Wow. So I would say you know, I think this might be a problem that's already getting solved, but an immense amount of GPU in the space would be. Can I actually? Can I say two things? We will make an exception. Yes, well, one I would. I would be two things in parallel. One would be just an unlimited amount of computing power and two would be physicians who had also been trained in data science and applied mathematics.

Ivanna Rosendal: 32:59

I can see how that would be a game changer. Yes, then we can start talking about like true augmentation of their capabilities if they understand how to interact with data. Those are great wishes. Well, irene, if our listeners want to learn more about Dainia Health or about yourself, where can they find you?

Eirini Schlosser: 33:24

Great. So it's on LinkedIn, irene Schlosser, or on our website, it's wwwdianiahealthcom.

Ivanna Rosendal: 33:39

And you can also find us on info at dianiahealthcom for emails. Oh awesome, this has been a lovely conversation. Thank you so much for coming on the show. Thank you for having me. You're listening to Transformation in Trials. If you have a suggestion for a guest for our show, reach out to Sam Parnell or Ivana Rosendahl on LinkedIn. You can find more episodes on Apple Podcasts, spotify, google Podcasts or in any other player. Remember to subscribe and get the episodes hot off the editor.