When asked how much of a role technology and tactics infused with artificial intelligence (AI) should play in the clinical-trial mix, Joseph Kim, senior adviser of patient experience and design innovation at Eli Lilly, can barely stifle a chuckle.
“That’s almost like asking, ‘Where should electricity fit in the mix?’” he says.
Kim’s opinion is all but universal within organizations that either conduct clinical research or work to coordinate, classify or contextualize it. Yet even as AI has replaced “patient-centricity” and “going beyond the pill” as the current go-to catchphrase of healthcare marketers, the industry has been slow to integrate techniques grouped under the AI umbrella into its research planning.
“In clinical trials, we are very cautious about incorporating new technologies because the risks of failure are so great,” says Laura Whitmore, director of applied innovation and process improvement at Otsuka Pharmaceutical Development and Commercialization. “I would suggest, though, that a new, possibly greater risk is emerging: failing to incorporate new ways of doing things. That might put us at risk of not enrolling people into clinical trials because patients will find our trials too cumbersome or inconvenient compared to how they live their lives.”
Oodaye Shukla, chief data and analytics officer for HVH Precision Analytics, puts a fine point on it: “Clients are realizing the old ways are not necessarily bearing fruit.”
To be sure, a learning/knowledge gap has something to do with the discomfort felt by many old-ways types. It doesn’t help that “AI” is alternately overused and misused when attempting to characterize these new techniques. “AI is a very general term. I don’t use it personally with clients because it’s so vague. It could mean 10 things,” says Andy Boothe, W2O’s head of data science.
For the sake of clarity, let’s define AI as any simulation of human behavioral process by a machine. Processes that can be grouped beneath the AI banner include machine learning (through which non-human systems can learn and improve without explicit programming) and natural language processing (through which computers are programmed to digest and analyze huge volumes of actual language).
In clinical trials, we are very cautious about incorporating new technologies because the risks of failure are so greatLaura Whitmore, Otsuka Pharmaceutical Development and Commercialization
Most people involved in the planning and organizing of clinical trials believe the industry is “directionally correct” in how it uses these technologies to render the process more participant-centric, according to Craig Lipset, formerly head of clinical innovation at Pfizer’s global product development group.
“In the past, humans were considered subjects, and subjects are very different from participants,” Lipset says. “When you think ‘subject,’ you think of a bowl of fruit an artist is painting. With all the tools we have now, we better understand trial design and the value and impact of engaging with patients early in the R&D process, not just when we’re desperate for enrollment and recruitment is struggling.”
Opportunities to use technology and especially advanced analytics abound. Robotic process automation is transforming the back-office mechanics of trial management. That’s a significant step forward because it allows data to be processed more quickly and for safety-related issues to be flagged earlier. Machine-learning techniques have also proven useful in flagging non-HIPAA-compliant data.
But the truth is the AI adoption rate in clinical trials lags that of other fields because of the industry’s traditional aversion to change, as well as the difficulty of shifting course in the middle of ongoing trials. “Sometimes it’s hard to move from these interesting and good experiments to doing it at scale,” says Lisa Shipley, VP of global digital analytics and technologies at Merck.
Indeed, new tactics and technologies tend to be added to existing processes rather than replacing them, which has the unfortunate effect of adding complexity to already intricate and costly schemes. “These tools end up being thought of as, ‘I have to do it the old way and the new way.’ Nobody’s willing to burn the boats,” Lipset says.
There’s no compelling reason why more — and more advanced — analytics-related techniques can’t find their way into clinical research. The most immediate impact could be felt in the maddening task of patient matching. When would-be participants volunteer for a study and are turned down, whether because they’re a poor match or they screen-fail, they often never step forward again. AI-type techniques can be used to ensure that participants are properly matched the first time out.
Natural language processing should be a no-brainer in situations such as this. It can help us understand and design research, then match research to patients.Joseph Kim, Eli Lilly
Then there’s the rich potential for natural language processing to make computer-human interactions more closely resemble human-human ones. Earlier this decade, a Lilly team took a deep look at the ClinicalTrials.gov website, which shows trial participation opportunities for interested patients. When the team analyzed how trial inclusion-exclusion criteria were written, it found the listings had, literally, 700 different ways of saying that would-be participants could not be pregnant during participation.
“Natural language processing should be a no-brainer in situations such as this. It can help us understand and design research, then match research to patients,” says Kim. “We need to make information more liquid and intelligible.”
He adds that technology has finally caught up to that need. “There are companies out there that can take what’s on ClinicalTrials.gov and make it ‘relate’ to clinical notes within EHRs from hospital systems,” Kim says. “That was always the barrier for matching patients to research: The two matching halves weren’t matchable. There was one type of language here and another there.”
Individuals pushing for more tech-savvy clinical trials will have to curb their enthusiasm and manage expectations accordingly. While the industry hopes to nudge participants toward pre-enrollment interactions executed by chatbots, participants aren’t all that keen on the notion: participants draw a distinction between chatbotting about the status of their J.Crew order and chatbotting about putting a drug in their bodies. That doesn’t mean AI-informed chatbots won’t someday figure into the trial mix, but it’s not going to happen as fast as advocates of it would like.
On the plus side, the industry is well on the way to getting its data house in order. Merck and Accenture recently announced a partnership with Amazon Web Services on a research platform designed specifically for life-sciences companies. If additional pharma companies sign on to the project — Shipley says Merck is meeting with other A-listers to gauge their interest — the industry could find itself on the fast track toward more uniform data management.
“If we’re going to take advantage of the data we have, common standards are so important,” Shipley says. “For AI to be used to its full potential, you need data in a format that it can truly be mined.”
As for the types of trials where AI-related techniques are a good fit, it’s worth remembering there is a very large difference between, say, diabetes research (which involves blood draws) and Alzheimer’s research (which involves answering hundreds of questions).
“It’s the conditions for which there are around 10,000 patients in the world where machine learning can give you some leverage,” Boothe says. “In a group that size, there are people who are undiagnosed or think they have some other condition. You can focus on certain language associated with the type of research you want to do — ‘I have a headache’ — and go from there.”
That’s why Shukla predicts technological innovation in and around clinical research will be fueled by small biotechs, which lack the resources to run traditional bells-and-whistles trials on their own. “If you talk to them, you hear things like, ‘I have to recruit patients and there’s lots of competition for them. How do I execute?’” he says. “That’s the kind of situation where companies are looking at machine learning as a kind of accelerant. That’s where we’ll see change.”