TORONTO—While the FDA has already doled out green lights to about 950 medical devices backed by artificial intelligence or machine learning, it has yet to approve a product that is capable of updating itself over time—one with the goal of becoming stronger as it meets more patients.
But that day is coming. So how are regulators planning to keep up with software that moves at the speed of AI?
“When you start thinking about the challenges, you first have to start by asking what's our biggest challenge,” said Troy Tazbaz, head of the FDA’s Digital Center of Excellence, in a panel discussion at Advamed’s MedTech Conference in Toronto this week.
“This session is about regulation that evolves, which is kind of an interesting way of putting it, because regulation unfortunately does not evolve,” Tazbaz said. “The regulation that we are using has been around since 1976—and essentially it was written for a very, very different type of product than what we're trying to apply it to. So a lot of our creativity has been asking how, with our current statutory authority, can we push the limits of that?”
One of the key hills to climb is ensuring that, as an AI algorithm gathers more data and refines its decision-making processes, it doesn’t stray from the levels of accuracy it first promised.
“Drift happens,” said Marc Lamoureux, manager of Health Canada’s digital health division. “Performance degradation happens. This is a fact. Obviously it depends on the model itself, and some can deteriorate faster than others, but performance today isn't necessarily a promise of performance tomorrow.”
Regulators aren’t necessarily facing a completely unexplored frontier. With the first AI- or ML-powered device receiving an FDA clearance in 1995—and many more in the years since, primarily in the field of diagnostic imaging and analyzing scans—agencies have realized the promise of the technology, even if they have only considered products with programs that are set in stone once they head to the market.
“The initial approvals were, not surprisingly, quite lengthy and complex—because it was new,” said Diane Johnson, senior director of digital health and U.S. policy for Johnson & Johnson. “But, since 1995, this has become commonplace. And while we're still in the locked-algorithm paradigm, for the most part, things are starting to accelerate.”
“But we don't have another 20 years to figure out generative AI and unlocked algorithms, like we did before,” Johnson said. “And at the end of the day, it is still always about patient safety.”
The clock is also ticking because many countries are preparing to care for rapidly aging populations, and AI is being counted on to help assist a healthcare workforce already facing burnout and shortages.
“In the U.S. by the year 2030, every single person in the baby boomer generation will be either retired or retirement-eligible, which means that they will also be Medicare-eligible,” said Tazbaz. “At the same time—and we have 22 million people working in the medical industry right now—by 2030, about 6.5 million are projected to retire, and we're only replacing them with about 2 million.”
“And what always happens when there’s an imbalance in supply and demand? Inflation,” he said. “Both in the cost of delivering healthcare on a per capita basis … and on top of that you're going to have delays in care.”
When it comes to regulatory approaches, the ideas behind predetermined change control plans—one of the FDA’s main pitches, with its draft guidance earlier this year, on how the industry can submit AI methods for review—have essentially existed in some form or another for years due to iteration being a pillar of the medical device innovation life cycle.
Health Canada, meanwhile, plans to publish a finalized version of its guidance on ML-enabled medical devices within the coming months, according to Lamoureux—who also said that transparency in AI development and describing how the models work will be key to building trust among providers.
“What we're seeing aren’t necessarily bad products—we're seeing good products that aren't being adopted at scale,” he said. “So what we want to do as a regulator is encourage the ecosystem to do what will make sense and build up that trust. Because we do see a ton of products, but at least in Canada, not a ton of adoption—there's a bit of apprehensiveness.”
There is also a question of who will shepherd the AI once it has been deployed, where different users may be feeding the algorithms completely different sets of data and ultimately drive the program toward different conclusions.
“The problem with the regulation in its existing form is that it doesn't factor in the proliferation of these models into 20 separate hospital systems,” said Tazbaz. “Regulatory agencies also tend to focus on things like adverse events. So what happens if there's an adverse event in one location, but the 19 others are just fine? What do you do with that product?”
But, from the industry’s point of view, quality assurance and post-market monitoring aren’t new, either.
“With traditional medical devices, deploying products to multiple hospitals is something we’ve been dealing with for a very long time,” said Johnson. “When you launch a new joint replacement implant, you get your approvals, you deploy it through all these hospitals, but the training—in this case, the physician, and not the algorithm—is the key to success.”
“Initially, you may have trained very senior physicians on how to use it, but when you move into people who are just coming out of medical school, how do you keep up the performance of the product?” she said.
“So if we are seeing different outcomes with an implant across surgical sites, then you have the option to go back and figure out where retraining may be necessary—or figure out if you have to mess with the design of your instrument. That’s something we've been doing for a long time … but we need to make it work fast enough to keep up with what's happening with AI devices.”
“These are not the challenges that existed back in 1976,” said Tazbaz. “The question for us now isn’t just what we can do with our current statutory authorities, but it’s about the way that we should be thinking about this concept of a medical device. I think that's what's evolving.”
“I believe all of this is going to push the boundaries of how we're thinking about this concept of a medical device, as more and more software gets embedded into our traditional hardware-based technologies," Tazbaz noted.