Artificial intelligence and machine learning have gained popularity in the medical device industry in recent years, with some top players in the space developing systems or buying their way into the competition.
Medtronic, GE and Philips have all invested in AI and machine learning, with claims that the technologies can better diagnose and treat patients. The wave of AI adoption and usage has brought new challenges to the healthcare industry, leading the FDA to consider adopting new regulatory review processes specifically for AI and machine learning technologies.
Rich Whitney, CEO of Radiology Partners, a U.S-based radiology practice, said that while interest in AI has been growing recently, barriers in the healthcare system like fragmentation and high amounts of regulation have still prevented widespread adoption.
One barrier is low adoption among physicians. However, Whitney said that Radiology Partner's recent partnership with AI medical imaging company Aidoc is intended to help address the problem.
"You need to make sure that physicians are on board and are properly trained and are really champions of this technology. Otherwise, it's not going to work. It's not going to be utilized; it's not going to get you the results you'd expect," Whitney said.
Nina Kottler, associate CMO for clinical AI with Radiology Partners, contends a primary benefit of using AI systems in radiology is improving patient care by identifying health concerns more quickly. Kottler said one of the crucial features of Aidoc's algorithms is a triage system, which can flag critical exam results for radiologists to prioritize.
"When your patient has an intracranial hemorrhage or a pulmonary embolist, these are findings that if you're not detecting them ... they can have catastrophic outcomes," Kottler said. "The earlier you get to those things, the better it is for patient outcomes."
The algorithms are also used for oncology exams, where they help identify if diseases are getting better or worse, according to Kottler.
Watchdogs, though, worry the pendulum could swing too quickly in the direction of AI. ECRI, for example, warns the technologies may be unreliable and misrepresent some patient populations, which could lead to misdiagnoses and inappropriate care decisions.
In a conversation with MedTech Dive, Whitney, Kottler and Aidoc CEO Elad Walach discuss how interest in AI has grown, handling potential bias and regulatory changes at the FDA.
This interview has been edited for clarity and brevity.
MEDTECH DIVE: How have you seen AI technology change over the last several years as this space has received more attention?
RICH WHITNEY: The technology is moving very, very rapidly. But we haven't yet crossed into that part of the evolution here where there is a significant amount of use and actual impact. The partnership with Aidoc really creates the prospect for much more widespread use of AI and really moving us into the future that we all envision, which is radiologists being enabled by AI and being able to add significantly more value to the health system.
NINA KOTTLER: There's been a lot of improvement in the technology, and a lot more options in terms of what kind of algorithms are available. But the technology on its own is insufficient. And I think what has been missing has been that connection with the radiologists. The technology is meant to be deployed in a clinical environment, and because there hasn't been a lot of deployment, there hasn’t been a lot of lessons in how to do that right.
AI systems have to be deployed with the direct assistance of radiologists to make sure that they understand how these clinical systems work. We need to make sure it's integrated into their workflow, and then we need to figure out how to monitor these systems over time to make sure that both the AI and the clinician are working together to improve patient care. And that's not simple.
Do you see providers prioritizing and investing more in AI systems today than two or three years ago?
ELAD WALACH: I can definitely say yes. By the way, COVID — even though it's difficult — really impacted the trend of healthcare executives being able to see value and return on investment from software-based solutions. They know that there is value to be captured by utilizing the right technical infrastructure and software. So I think that in terms of prioritization, absolutely. But I think there is a lot of momentum building up by physicians and radiologists using the technology, understanding that there is value and analyzing what that value is.
The FDA is considering how best to regulate AI. For example, whether to allow algorithms to be updated without review or remain "locked." How will a change in this review process impact the industry?
KOTTLER: Instead of just locking an algorithm in at one point in time, and then waiting for that algorithm to improve and redoing that evaluation, the FDA is looking at evaluating the vendor and their practices to see if the way that the vendor updates things themselves is good enough. And if the vendor’s processes are good enough, then the output should be good enough. So, the agency won’t actually have to check the output, they can check the vendor processes. They're just in the very beginning of it. I think they're beta testing it with a few big groups right now, so it's going to take a little while. But I think it's quite fascinating.
WALACH: It is a difficult problem the FDA is facing, and a lot of it is the flood with the number of products that are coming to market. The question that the agency is tasked with is, How do we maintain safety and efficacy while making sure that we can bring innovation to the market? The agency has been moving quite quickly in terms of creating new processes, new pathways and being very communicative with the companies. So, you've asked, Are you waiting for the FDA to do something? In some sense, yes, but it's a very active process. It’s an active engagement with the agency. I do think that there are some exciting regulatory changes ahead.
One issue continually brought up with AI is bias built into algorithms. How do you work to prevent this from happening and then fix the problem if it is recognized?
WALACH: You want to make sure that on the one hand, you trained the data on a very robust, diverse set that isn't biased towards a certain population. On the other hand, we want to make sure that even after we release a product to market, we may encounter bias that was unexpected initially. We want to make sure that we keep monitoring performance over time. For me, it's battling biases with data. That data is the protector against bias in all stages of the product lifecycle.
KOTTLER: Eventually, we may end up going in the opposite direction. Right now, we're trying to have data that's as generalizable as possible so you can apply the same algorithm everywhere. But, ultimately, that means that the specificity and value for a certain patient will have to decrease, even if it's just a little bit.
As the FDA evolves, and as these AI algorithms evolve, we will be able to have an AI algorithm that's suited for a specific population, and that means it's going to be much more accurate for that population.
Where do you see AI use heading?
KOTTLER: The next area that we're getting into is predictive medicine. While medicine has always been about the treatment of disease, we really need to move more into the prevention of disease. AI can help us with the prevention of disease because it's detecting things that we may not be able to detect as humans. If we combine that information with the other information that we have as humans, we can start to predict which patients are more at risk.
For example, for breast cancer, maybe certain patients should be getting mammograms or their imaging studies much more frequently than others. Maybe we can identify if certain patients are at risk for developing a bone fracture because we can look at the quality of their bones and see which ones are the most at risk for developing osteoporosis. These are all preventative measures that I think we're going to get much more involved in.
We're going to combine that with information from the patient systems that are getting more prevalent, like wearables, to provide a more holistic view of the patient.