A number of entities have sounded off on the FDA’s discussion paper for artificial intelligence (AI), including several medical societies that would like to see autonomously operating algorithms subjected to more stringent review than supervised algorithms. Vibhor Rastogi, general partner at Symphony AI of Los Altos, Calif., told BioWorld that the company is on board with many of these concerns, adding that the FDA discussion paper does a “good job of balancing innovation and patient safety.”
Among the organizations that offered their perspectives on the FDA’s AI discussion paper were the American College of Radiology (ACR) and the Radiological Society of North America (RSNA), which said they had misgivings regarding algorithms that would operate independently of close physician oversight. One of the concerns expressed in the joint medical society letter was that some of the more limited data sets used to construct algorithms might lead to poor performance across heterogeneous populations, while an industry group indicated some concern that the data demands described in the paper could ultimately amount to little more than a useless data dump.
FDA seen as appropriately balancing innovation, safety
Rastogi, who has invested in several AI efforts in the past couple of years, said the joint ACR/RSNA letter is well thought through and articulated the most important concerns about AI and what algorithms need to address for adoption. “A number of investments we have made align with the concerns” voiced by the medical societies, he said, including the need for generalizability of the data, the ability of the algorithm to work properly across different conditions, and the need to have experts in the loop before concluding that the algorithm has presented an accurate diagnosis.
“I do believe that FDA, in their regulatory role, has done a good job balancing innovation and patient safety,” Rastogi said, adding that he believes the FDA will steer the appropriate regulatory course, including that the agency will not permit physicians to be cut completely out of the diagnostic loop.
AI is being deployed more or less rapidly in several areas of the economy, such as in e-commerce. “As investors, we have embraced the view that clinical algorithms do need to be handled differently,” Rastogi said. Symphony is working with algorithms in several domains, including stroke, and the company is interested in seeing a regulatory framework that mandates a change protocol for autonomously adaptive algorithms and post-market monitoring of real-world performance. “We incorporate that into our release cycles” already, Rastogi said.
The appropriate course for the FDA would be to ensure that surveillance of an algorithm in real-world usage is based on an established methodology, but Rastogi said he is confident that the agency won’t let these algorithms onto the market without proper validation, either. “FDA is taking the right approach,” Rastogi said, adding, “these can be life-and-death decisions, so they have to follow change control protocols.”
One of the fundamentals for AI is that the algorithm has to be explainable to the medical professional and the results must also be verifiable, Rastogi said, adding that the radiologist has to be able to understand what the input parameters were and how the algorithm modeled those inputs. An algorithm “has to have that full end-to-end visibility, and has to be highly interactive,” he said, to the degree that the radiologist could tweak some of the features and change the output to ensure that output is clinically justifiable.
While such a requirement might initially come across as onerous, “we think that is not a rate-limiting factor because for the foreseeable future, at least in radiology,” AI and humans “will have to work together,” he said. Even with this kind of transparency, “we don’t think innovation in radiology will be hampered” or that regulations will become a rate-limiting factor for adoption, Rastogi said.
Class III designations unlikely for most AI
There is no particular reason to believe that most algorithms will receive a class III designation, given the indications for use for most of the algorithms now on the market, Rastogi said. This assumes the expert physician or radiologist will review the results, but a developer can help ease any fears on the part of the FDA and end users by incorporating a robust and diverse set of data points to develop and validate the algorithm.
As for the cost of obtaining those data sets, Rastogi noted that not all hospitals are particularly aggressive in seeking to monetize their data. Some who run these hospitals would like to help push the clinical availability of an algorithm, and while Symphony AI is not averse to reimbursing hospitals for their data, these data sets are not as difficult to obtain as is sometimes presumed.
Rastogi said he is unaware of any reason to believe the FDA will need new statutory authorities to deploy a novel AI regulatory mechanism, which reflects the position taken by Scott Gottlieb when he was the FDA commissioner. Rastogi also said that Gottlieb “brought a lot of that investor and market perspective to the FDA,” a perspective that has sustained at the agency since his departure from the agency in March 2019.
The twin questions of coverage and reimbursement are often as complicated as regulatory review, and Rastogi indicated that this is true of AI and machine learning as well. “We do think there needs to be more reimbursement for these algorithms,” he said, noting that many are covered under the same professional and technical reimbursement codes the radiologist would receive without the use of the AI.
The new technology pass-through mechanism does not seem to have served as a particularly helpful vehicle for additional rates despite that these algorithms add a lot of value to the standard of care, Rastogi observed. Should the Centers for Medicare and Medicaid Services devise a novel reimbursement mechanism to fill in these gaps, “that would be a very strong impetus for industry, similar to how reimbursement drove growth for 3D advanced visualization” for applications such as angiography, he said.