From discovering compounds that could eventually become drugs to training chatbots to support patients taking those drugs, artificial intelligence can play a role anywhere along the drug development cycle. One of those points is clinical trials, where this technology can make studies easier for patients and investigators to navigate, and ultimately help biopharma companies make the best drugs they can.
The key is selecting the right tool for the job, said panelists during a Fierce AI Week discussion. “Artificial intelligence” in biopharma has become an umbrella term encompassing true AI, machine learning, deep learning, computer vision, biostatistics and data science.
“For us at Janssen R&D, it really starts with the question when we look at the pipeline, the different medicines we are trying to build,” said Najat Khan, Ph.D., Chief Data Science Officer and Global Head of Strategy and Operations, at Johnson & Johnson’s Janssen.
“Once you have the question identified that’s pertinent for the program… then we say, what are all the different approaches we can use to actually answer that question?” she added.
Some of those questions include how sponsors and investigators can make sure they’re using the right endpoints, or how studies can track disease progression, especially in diseases that may present in different ways in different people. In the latter case, study teams could be juggling various datasets, including genomic and imaging data to physician notes–and they’d need the right tool to fit the task.
“It’s a broad range of capabilities; it’s not always the most sophisticated algorithm that will answer your question, but what is fit-for-purpose for the question you’re trying to answer,” Khan said.
But that’s just one piece of the puzzle. Another part is how to build and train AI models so that they’re useful and not causing harm, said James Hamrick, senior medical director at Flatiron Health.
“We are in many ways training a machine to simulate human intelligence, so the processes of learning and reasoning, as well as self-correction,” said Raolat Abdoulai, M.D., clinical research director at Sanofi.
“We need to make sure we are providing it with appropriate information and data for it to learn from… either from past trials as well as data in the real world in order to make better, effective and more reasoned decisions,” Abdoulai added.
Flatiron, for instance, has models that can predict whether a patient has metastatic disease, or if a patient is likely to be admitted to the emergency room during cancer treatment.
“At the same time, we are retrospectively running a check on the model, looking at different categories of patients–so things like gender or ethnicity–and we are actively checking the model as we go to make sure it is in agreement and bias-free,” Hamrick said. “And what I mean by that is, the predictions we made two months ago–what happened in those two months? Were they accurate across these different subgroups of patients?”
While drugmakers should make sure their models work for diverse patient populations, they may use other AI tools to ensure they enroll diverse patient populations in the first place, so there aren’t any nasty surprises when it rolls out the drug.
“We’ve done that in some trials where we know that the disease impacts more diverse patients and our clinical trial participants should be reflective of the real-world setting… Instead of going to a site and asking, can you have this many diverse patients, you flip it around and say where are the diverse patients to begin with?” Khan said.
“In this way, we take away the bias that prevents patients from a diverse range of backgrounds from being recruited,” Abdoulai said.
Beyond improving data analysis for drug development, AI tools, mobile or wearable devices and remote monitoring could also make open up clinical trials to people who perhaps don’t live close to a trial site or may not want to visit a site, especially during a pandemic. These tools could be the key to not just digitizing trials, but also decentralizing them.
One plus would be bringing “the technology to the individual, the subject being recruited, to where they are and rather than bringing them to the clinic and subjecting them to analog, non-digital ways of acquiring data,” said Ajay Royyuru, IBM fellow and vice president of healthcare and life sciences at IBM.
“We have digital health and ways of acquiring data remotely that actually lend themselves better to not only fairness, but also to AI-based analysis to continuously acquired data. This is beginning to occur, of course. Many digital biomarkers and digital endpoints are getting defined,” Royyuru added.