How Syneos Health is Building AI Solutions for Clinical Trials

By Daniel Pereira |  March 25, 2024

Syneos Health, based in Morrisville, N.C., offers support services to biopharma companies as they test new drugs in clinical trials and later launch them in the market. The publicly-traded company reported annual revenue of $5.4 billion in 2023, and operates in 110 countries with nearly 30,000 employees.

Right now, says Ilya Vedrashko, Vice President of AI Products,  the focus when it comes to creating and deploying new AI software “is on creating the solutions for Syneos [employees], and upgrading a lot of the internal processes with the solutions. Our headcount alone creates enough demand for us to justify building our own software.” Eventually, the company may offer some of the software it develops to client companies.

In a recent conversation with Vedrashko, he talked about taking a “bottom up, problem solving approach” to understanding where AI and machine learning can be best applied. This bonus interview was conducted as part of our recent research initiative, “How AI is Influencing Corporate Innovation in 2024.”

• • •

Ilya Vedrashko, Vice President of AI Products, Syneos Health

What can you reveal about the most bleeding edge thing you are doing with your AI platform?

I can talk about the space that keeps me up at night and interested enough to work on it after hours on my own time — the clinical trial space. Clinical trials have certain standard elements. There are certain milestones, certain types of roles, standards, procedures. One of those is the clinical study protocol which describes how the trial needs to be conducted, what procedures need to be administered, what assessments need to be taken, in what order, how it is all calculated. 

Those protocols exist in a text format, usually in PDFs or sometimes text descriptions of the clinical trials on a government website or similar regulatory websites. There is a lot of knowledge that can be mined from the data. Mining it and converting it into a structured data set is something that me and my team have been working on for over a year. And it shows a lot of promise. It’s a difficult problem because it’s a specialized space. The tools that exist are not built specifically for this purpose. So, we’re building a lot of our own knowledge and methods so that we eventually produce a solution for ourselves and a future product. 

Can you share a use case you’re focusing on?

I will give you one solution on which we are further ahead. There is a concept in our space called “patient burden” — how demanding is the clinical protocol or the clinical trial to the patient? …There are protocols with complex conditions, with complex treatments that potentially unfold over multiple years, and that places various kinds of requirements on the patient. Measuring those requirements, or “burden,” is something that is important for the industry — obviously to be able to minimize that burden to make sure that the patient is inconvenienced as little as possible…

Doing it in a way that is structured and repeatable has been challenging because a lot of the information that is used to determine that burden has been locked into that unstructured data — in those PDF documents or text-based descriptions. So building the bridge between understanding and structuring the protocol text to understanding and creating a mechanism that would allow us to compare the burdens placed on the patients by the different trials. That is something that is worth working towards and something that we’re making satisfactory progress on already.

…We evaluate models and products and determine if they can help us or not. If not, then it just becomes an extracurricular activity.

Because you lead an AI product team that pre-dates the release of ChatGPT, I have to ask: how have you and your team reacted to the hype cycle around large-language models (LLMs) and Generative AI?

We’re human. There’s been some eye rolls, of course. But I think grounding ourselves in what we need to ship helps us keep the attitude very pragmatic. So we evaluate models and products and determine if they can help us or not. If not, then it just becomes an extracurricular activity. And if it does, then we spend some time figuring it out and understanding it. And that means understanding the technology, playing with it, setting up infrastructure, and also engaging with other folks outside of our company who are doing the same.

Does each business division engage with your AI solutions? Do they need a certain tech skill-set?

Learn more about this report, published in February 2024.

We have a community of people who are comfortable working with code and with the data directly. So we are creating a data repository internally that folks with those skills and those tools can access and build their own things. For the business users who are building applications, they have a GUI-based tool where they can go, formulate the problem, and get the outputs without having to write their own code from scratch. So, it’s kind of a dual-purpose structure. It is open source within our ecosystem. So prompts and the prompt libraries — people who are interested in doing large language models (LLM) are able to work on their own to evaluate them. They have data, codes, models and have all of this registered. We have pretty robust documentation going on, so whoever is interesting can pick it up and do something with it.

How are you framing AI experimentation or exploration for non-technical persons, or how do you encourage business leaders to get started or plug into your team and the tools you make available within the company? 

You can always distill any course of study to like one or two things that you remember, and then everything else is just derivative. I think for me, it has been the focus on people – and this whole idea of humanism as the center of any technological innovation. And that is such a simple but mind-bending concept. But it helped arrange a lot of the other things in my head, and also formulate how we behave. It has also proven to be especially valuable in this industry vertical. Because once you put people at the center, it also actually works with how the company thinks about its work – again, concepts like “patient burden.”

For Syneos, it is all focused on the patients – and once you focus on the patients, then everything else [arranges itself] in a certain way.  It is the same here. Once you focus yourself on the user, on the business user, the intended user of the applications, then everything else arranges itself accordingly.  That means that your language needs to adjust.

Product design thinking doesn’t only apply to software, it applies to your language, it applies to your emails, it applies to your PowerPoint presentations.  So you create a positive user experience for everything you produce. 

Are you finding there is already a large language model (LLM) and Generative AI sub-sector specific to your sector?

There is definitely a lot of interest from the leaders in this space. Google and Microsoft build solutions specifically for the healthcare sector. A lot of the data in the healthcare space is textual, but complicated textual… so it lends itself well to the kind of processes that LLMs enable, but they also require enough specialization to be useful and to reduce the error rate. Such efforts require significant investment. So we do see a fair amount of effort happening to create the infrastructure that is specific to the healthcare space — which we will also be able to use.

So that is from the provider side. And then on the demand side, we see our colleagues, in grassroots fashion, experimenting with different solutions to their problems, but also our clients trying to solve a lot of different problems with pieces of the technology. So I don’t think I have seen a Hugging Face-type marketplace emerge yet due to these unique issues  – and that overall complexity. But we do see a lot of efforts that eventually are going to converge somewhere.

Implementation is the tough part and the boring part nobody likes to talk about.

What is one piece of advice you commonly share with other innovators in big companies?

Innovation has two parts, invention and implementation… Invention is the more exciting part, it seems. That is where you get to do all of the sticky notes, all of the workshops, and the brainstorms. Implementation is the tough part and the boring part nobody likes to talk about. I love the implementation part too — probably even more than invention. I’ll be pretty happy with doing the old stuff — but make sure that it is done right, and it is very process oriented. So that is what I share with my team: all of the innovative ideas, you have to think them through to the end. Imagine them in the wild, right? And that is a great filter.

Is the message “learn to love implementation”?

Yes. Otherwise, if you don’t follow through, all you’re left with is innovation theater, which has its place…but you are not unlocking the entirety of the economic value if you don’t follow through in the implementation phase.