Parexel 首席患者官 Stacy Hurt：AI 设计临床试验时可能漏掉什么

This audio is auto-generated. Please let us know if you have feedback.

When Stacy Hurt looks at a clinical trial protocol, she’s always thinking about something beyond its endpoints — her 21-year-old son Emmett. Diagnosed with a chromosome abnormality so rare that his syndrome doesn’t have a name, Emmett depends on his parents around the clock to meet his basic needs.

“I look at clinical trials as if my son were participating in them because I'm very hopeful that cell and gene therapy will improve his life one day,” Hurt, chief patient officer at the contract research organization Parexel, said.

That human-focused viewpoint is colliding with the growing presence of AI, which is playing a greater role in protocol design as trial complexity rises and average enrollment drops.

Other personal experiences have also made Hurt acutely aware of where the technology excels and where it falls short. As a stage four cancer survivor, Hurt is leveraging her unique perspective and role at Parexel to help ensure patients are in the technological loop at the earliest stages of clinical design, even with more AI in the mix.

“When we have a protocol, we take it and we internally run it through some proprietary AI tools to optimize it,” Hurt said. “As chief patient officer, I will put in prompts [like]: How can we reduce patient and caregiver burden in this protocol?”

Some protocols require multiple blood draws, Hurt pointed out, noting that this kind of situation would be hard for a patient like Emmett to tolerate. The same is true for protocols requiring multiple appointments or long wait times. Other patients may have different challenges, including financial or logistical barriers.

But making sure the patient voice is prominent in trial design is becoming more critical as the industry increasingly leverages AI. 

New tool, new challenges

Companies like Bristol Myers Squibb are using AI in hopes of accelerating trial planning, according to a report by Intuition Labs. The technology can provide insights on design consistency, flag scheduling conflicts and identify potential inefficiencies. AI may also help identify trial participants, according to a McKinsey & Company analysis, which found that AI and machine learning tools used in clinical trial operations could boost enrollment by 10% to 20%.

But while AI excels at nuts-and-bolts tasks, it can lack the ability to fully understand the treatment experience, even when it’s trying to account for the patient experience.

In a review of one oncology trial protocol, for example, Hurt said the AI tool flagged the amount of time a patient would spend in the chair receiving the study medication as a trial burden. But it failed to account for the fact that patients undergoing the experimental treatment would have spent the same amount of time receiving standard chemotherapy, she said. Hurt immediately recognized the AI’s inadequate logic, having gone through 55 rounds of chemotherapy herself.

Patients also ask questions that others might not.

Hurt, who has been cancer-free for more than a decade after undergoing a grueling multi-year treatment, spoke up at one pre-trial workshop with a sponsor developing a colorectal trial protocol to find out if patients with liver stents would be excluded from the study.

“I talk to advanced colorectal cancer patients all the time, and that's the biggest exclusion criteria,” she said.

The sponsor remarked that the question was one that only someone with lived experience might know to ask, Hurt added.

And while AI may be adept at choosing scientific endpoints, other measures matter too.

“We point to overall survival in oncology studies, and that's great,” she said. “But if a patient has had multiple surgeries, multiple interventions, and they are bedridden and can't get out of bed, what good is it?” 

The cost of bad design

Poorly designed clinical trials not only deter or unnecessarily burden patients, but drive up costs.

“One protocol amendment in a phase 3 trial can cost $535,000,” Hurt said, referring to figures from a 2016 Tufts Center for the Study of Drug Development analysis of the median direct costs of substantial protocol amendments. “In phase 2, it's $141,000.”

Including patients may help identify design problems before they require high-priced changes. The technology’s pattern-detecting abilities may also help flag patient safety risks or early signs that a patient may drop out. But technology still needs humans in the loop. 

Hurt’s concerns about AI don’t mean she doesn’t support the technology, she pointed out.

“I'm a huge proponent of AI,” Hurt noted. “I'm mostly passionate about how AI can help clinical trials fit into patients' lives, not the other way around.”