DIA 2024: Machine Learning and Simulations to Facilitate Clinical Trials

Image Credit: Andy Studna

In a session at the 2024 DIA Global Annual meeting in San Diego, CA, Raviv Pryluk, CEO & co-founder, PhaseV; Sam Miller, head of strategic consulting, Exploristics; and Andrew Stelzer, head of business development, Unlearn.ai addressed a number of ways simulations are currently being utilized in clinical research.

Miller began by highlighting how challenging the industry is, citing the long process of bringing a potential therapy from Phase I of a trial through to the approval stage. “Typically, only 10% of drugs that start with Phase I can get to the approval stage. That's enormous failure,” he said. “It's not changed very much at all in the decades that I've been involved in the industry. If anything, it's got worse, but something that has changed is our access to data.”

With so many recent advancements in technology, such as those that improve data access, how can industry get more approved therapies to patients?

“So, if we’ve got all this data—things that we could potentially learn from and improve—why aren't we doing it? I would say a reason is that we're a bit too slow to change,” Miller continued.

When it comes to creating data—data that is often missing from patients not completing the study—there are three important steps.

“The three steps are to create data that looks as similar as possible to the data you're going to collect in your study; it's got the outcomes, it's got the baseline measurements, it's got the interdependencies and correlations and structure and so on. Then once you've generated that data, you can sample from that and effectively create synthetic virtual clinical trials, in advance. You can analyze those sets of data, and you can quantify what proportion of those virtual studies are going to be successful, and that gives you a lot of information in order to make decisions about which are the key elements of your study design that you want to apply to maximize your chance of success,” Miller explained.

Stelzer then highlighted the process of creating digital twins using artificial intelligence (AI) and machine learning. He noted that industry often has misconceptions around what digital twins really are. While a digital twin is not data from a matched patient or historical data, it is a model to determine future outcomes. “A digital twin is not like a new patient at all,” he said. “It is a model-based forecast for the future clinical outcomes for a specific trial participant.”

To create these digital twins, two necessary components are training data and an AI model. Once the data is in place, then the AI model can be trained on creation. According to Stelzer, these digital twin generators are AI models that forecast time points using baseline and previously forecasted data.

“Our goal is to be able to have a digital twin generator or AI model for every disease that pharma and biotech companies are working in so that we can enable them to use digital twins,” Stelzer said. “So having digital twins is great. It's a technology at its core.”

Pryluk concluded the presentation by highlighting the challenges with AI and simulation in clinical research. While there are many that currently need to be addressed, the benefits are worth the hard work.

“There are many challenges that we need to overcome as a community, but once we are incorporating models that are generalizable, validated, explainable, actionable, and with statistical guarantees, we can start seeing how we are gaining the benefits that are unknowns,” Pryluk said. “I hope that in the next few years, we're going to see more and more machine learning tools that are empowering all of us to do faster, more ethical, and more successful clinical trials for patients.”

Reference

Pryluk R, Miller S, Stelzer A. Machine Learning and Simulations to Facilitate Clinical Trials. June 17, 2024. 2024 DIA Global Annual Meeting.