From The Editor | January 21, 2021

Is AI Improving Outcomes In Clinical Trials?

Ed Miseta

By Ed Miseta, Chief Editor, Clinical Leader
Follow Me On Twitter @EdClinical

AI Artificial Intelligence

In October 2020, I hosted a webinar on the topic of AI and how it will improve outcomes in clinical trials. Our panel featured:

  • Jennifer Goldsack, executive director at the Digital Medicine Society (DiMe)
  • Nechama Katan, director, data science lead at Pfizer
  • Alex Zhavoronkov, founder and CEO of Insilico Medicine
  • Elvin Thalund, director of industry strategy at Oracle Health Sciences

This article highlights some of the comments made during that discussion.

Ed Miseta: Elvin, I'd like to start with you. Clinical executives seem to have a lot of questions about AI. What would you tell them about how you expect AI and machine learning to impact clinical trials in the future?

Elvin Thalund: The first thing to keep in mind is that this is a new technology that will enable us to look into the future. That will be possible due to the predictive capabilities of the technology. And I think we need to broaden our perspective. We can talk about clinical trials, but we should be thinking more in terms of clinical development. Companies need to be able to look across their Phase 1, Phase 2, Phase 3, and even their Phase 4 trials and be able to predict what will happen. For that reason, I think companies need to look at the big perspective.

When we look across trials, I think the true benefit will come from avoiding failures. When the trial is up and running, your time is spent keeping the train on the track. Start your planning early, before you even start the trial, in the planning phase of your clinical development. That will provide the biggest benefit because you can plan ahead and take the best approach. For example, should you plan a decentralized trial? You need to think about that before you write your first protocol. This where AI and machine learning can help predict the right study design.

Miseta: Jennifer, I was hoping you can jump in on that first question as well. You talk to a lot of people. When you get questions about what will AI do for my clinical trials, how do you answer them?

Jennifer Goldsack: I feel like a bit of a brat when I answer that question. People like to ask, How am I going to apply this technology to my trial? In response to that, I like to ask, What is the problem you're trying to solve? The solution should tie into the problem. We can look at different technologies such as machine learning, neural networks, deep learning, or AI. They're all subtly different, although there is a hierarchy since some sit under others. I think there's also confusion that exists because we think more about buzzwords and being thoughtful with our language. As a result, I don't think there's a shared understanding of what these different terms mean. We need to look at the challenge first, not the technology. Elvin described some problems we can solve with machine learning. Understanding the problem you face is always the first step.

Miseta: When talking about AI, I hear many experts say what we have is not really a technology problem, it's a human understanding problem. Nechama, let me jump to you. Are we at a point right now where the folks who will be using the technology just don't understand how it will impact them and what they need to do to prepare for it?

Nechama Katan: This really has to do with the question you are asking. We have a lot of data. I work in a data management organization, but I could work in any other part of the clinical trial cycle. Regardless of where you are, we have a lot of data. We also have a lot of very manual processes in the industry. This is one area where AI will add value. But that can only happen if you ask the right questions, and asking the right questions requires users to understand the language of data science. To use machine learning and AI we need to understand the problems that need to be solved and which problems can be solved through the use of these technologies. We also have to frame these problems in a way that someone with access to the right tools and data sets can solve them. The technology is not the limiting factor right now.

Miseta: For those employees who do not understand the technology, how do we teach them what they need to know so they can start asking the right questions?

Thalund: I have participated in webinars on this topic. I believe this is still framed as a technology problem even though it isn’t. In my opinion, machine learning is simple. From a business point of view, you have to define what are the questions you are trying to answer. But answering those questions means you also have to understand your own business and what you want answered, because that’s what machine learning would predict. Machine learning will also start giving you real insights on what will make a difference and impact your goals. Once you know that, you can begin to collect the data.

Goldsack: This is an important tenant of a service DiMe performs. We recognize the digitization of health care will require extraordinary interdisciplinary knowledge. However, it does not require that every clinical executive or clinician go out and get a PhD in data science or in computer science. One of our archived educational sessions that is accessed most regularly deals with machine learning. For that session we bought in a set of experts from MIT, Evidation Health, Flatiron Health, and other organizations. In 45-minutes we describe what ‘good’ looks like and discuss the right questions to ask and what you should watch out for.

The education piece is important, but I also think we need to stop having silos between our clinical experts and our technical experts. Every clinical company does not have to become a technology company. However, I think we need to have a trusted technical member of the team who is not in another building and who is part of the strategic decision making. If the industry is serious about using a variety of different digital tools to solve some of their most pressing problems, we will be able to bring benefits to our patients. We need to have a higher baseline of understanding across all disciplines. For example, we expect technical colleagues to have a sense of clinical workflow. But we also need to eliminate the silos that exist between the different sets of expertise we need to bring together.

Miseta: Study startup is a challenge for most clinical executives. How will machine learning help companies with the challenges they face in that area?

Thalund: Today, I think approximately 50 percent of all trials end up under enrolling, and enrollment in many cases is being pushed through by brute force. And if you start looking at what it takes to predict what patients and sites are out there, then you have the leading indicators for machine learning in your clinical development plan. That enables you to predict enrollment, with a traditional trial design, before you write your first protocol. Today, we write a protocol and then find out after the fact that you cannot enroll the patients you’re expecting. With machine learning you can predict enrollment ahead of time, and then write the right protocol.

But, more importantly, you will be able to design protocols for what you need to achieve. A limiting factor for decentralized trials will be the ability to predict enrollment. We do not yet know how that design will work, and that unknown may cause some companies to fall back to traditional trial designs. That is where machine learning can step in and enable you to know if a model will work. That will help companies avoid failures.

Alex Zhavoronkov: I agree and can’t overemphasize the importance of the ability to predict outcomes, as that will enable companies to decide if they should proceed with a trial and how they should prioritize programs. The power of AI today is that ability to predict the future.  

I believe AI will allow us to train our models and then see how they perform. Some of those models can be used for enrollment and allow us to predict patient participation, but we can then use that same model to help us design the trial better. There is a lot of power there. AI can also help us determine which patients are more at risk for side effects, allowing us to exclude them from studies.

AI allows us to decipher those black boxes and really put those models to use. For prediction, enrollment, and even other simple rules. For example, as a predictor of biological age. Biological age can be used for clinical trial enrollment in an oncology study. People who are predicted to be significantly older than their chronological age are more likely to die prematurely from something other than the cancer you’re addressing. Therefore, those are folks you may not want to enroll. I see a breadth of applications for AI in clinical trials, and I think we are currently scratching the tip of the iceberg.