FDA leader talks evolving strategy for AI and machine learning validation
At a virtual meeting of the U.S. Food and Drug Administration's Center for Devices and Radiological Health and Patient Engagement Advisory Committee on Thursday, regulators offered updates and new discussion around medical devices and decision support powered by artificial intelligence.
One of the topics on the agenda was how to strike a balance between safety and innovation with algorithms getting smarter and better trained by the day.
In his discussion of AI and machine learning validation, Bakul Patel, director of the FDA's recently-launched Digital Health Center of Excellence, said he sees huge breakthroughs on the horizon.
"This new technology is going to help us get to a different place and a better place," said Patel. "You're seeing a great opportunity. You're seeing automated image diagnostics. We have seen some advanced prevention indicators. Data is becoming the new water. And AI is helping healthcare professionals and patients get more insights into how they can translate what we already knew in different silos into something that's useful."
As new tools like those are deployed to "augment what we already have in place," he said, "we're also seeing that evidence and information that used to be in different areas that were only locked up in places, technology and machine learning and algorithms and software is bringing that together and will help us get to a place where we are all better informed."
We're at a pivotal moment where "software can take inputs from many, many, many sources and generate those intentions for diagnosing, treating," said Patel.
"As we start getting into the world of machine learning and using data to program software and program technology, we are seeing this advent, and the fluidity and the availability of the data, becoming a big driver," he said. "And that comes with opportunities – and that comes with some challenges as well."
One of those is the fact that both the technology and the datasets are evolving at lightning speed.
"There's data sets required for supervised learning, unsupervised learning. And then when we start thinking about deep learning, where the machine learns about the inherent characteristics of the data, rather than looking for informed data," said Patel.
The good news? "As we start going further down in this technology pathway, you will probably see better and different techniques emerge as we start moving forward. The question that really excites us is how can this ability of machine learning algorithms and systems that are learning from the wealth of information that is available to them can potentially develop novel AI and ML devices – for all medical devices, for that matter."
Patel said FDA sees a future where AI "can start detecting diseases earlier, can accurately diagnose – and accurately rule out. Personalization is an aspect that we feel that can be empowered by machine learning."
That said, however, capitalizing on those advances depends on reaching a delicate balance between empowering innovators and protecting patients.
"Our goal has always been how do we enhance patients having access to these high-quality digital medical products?" said Patel. "How do you allow manufacturers, on the other hand, to rapidly modify?, because this technology is changing or over and over again, as we as humans, and the machines to learn, but at the same time maintaining reasonable assurance of safety and effectiveness, while it's trustworthy and minimally burdensome for all."
That's a model that's "evolving," he said.
"People are using data to train and tune a model, and validate it, and then putting it out into deployment. But the biggest change that's happening now is the machine itself can learn from users. And that input by itself is fed back into the model. As these machines learn, we feel like there is going to be a change in expectation."
Getting to the next level safely and efficiently is going to depend on "trust and transparency," said Patel – especially as the technologies get more and more advanced, ever more quickly
On one hand, "there's the space where things are learned and locked and where the products are deployed," he explained.
"But then on the other end of the spectrum, you can imagine these systems that can learn on an ongoing basis. And that could [be] every time the machine encounters a new situation, or could be much more frequent than that."
Even in a situation where the advances are coming fast and furious, however, "some of the foundational questions don't go away," said Patel. "When we are talking for medical purposes, we want to make sure that the valid clinical association exists. That there is a validation on our side, and the clinical validation exists that we can all trust."
The challenge then, as the industry moves apace "into this continuously learning world," he said, is what sort of mechanisms enable that innovation balance. "What does that framework look like?"
At organizations such as the International Medical Device Regulators Forum, for instance, there's ongoing work around forward-looking concepts to manage machine learning from real-world information that can be fed back into the system, he said.
Beyond that, however, there are more basic imperatives. The big one, of course, is that "the quality of the data is something we need to have assurance on," he said.
"We all know there are some constraints, because of location or the amount of information available, about the cleanliness of the data. That might drive inherent bias. We don't want to set up a system where we figure out, after the product is out in the market, that it is missing a certain type of population, or demographic, or other aspect that we have accidentally not realized."
But even with "large, high-quality curated data sets," said Patel. "There's also a need for users to understand what the machine is doing, what the software is doing, ... how the machine learns, what is learned, what is retained. That's going to be something we need to be clear on.
As things move forward, "one fundamental thing I would want to say is that we need some separation," he explained, "separating the training from the testing from the validation datasets are very commonly used in this space.
"We also want to make sure that things are consistently used in practice in this space," he added. "You want to make sure the learning process, and the testing, and the validation process is transparent to the users."
He noted the agency's total product life cycle approach to AI-powered software-as-a-medical device, "where FDA oversight would provide the level of trust and confidence to the users, at the same time leveraging transparency and pre-market assurance, as well as ongoing monitoring of those products that are learning on the fly. And we are looking to see what we can do to enhance this framework going forward, and understand how the regulatory system can enable that."