The advent of AI at the US Food and Drug Administration (FDA) is changing the complexion of the regulatory approval process for medical devices.
In May, FDA commissioner Dr Martin Makary announced that the agency would be implementing Elsa, a generative AI (genAI) tool intended to drive efficiencies at the agency. Elsa adds to an increasing range of templates and other automated tools that have infused the agency in recent years, with a specific focus on the review process for medical devices and drugs.
Elsa鈥檚 rollout ahead of schedule last month proceeded mass layoffs at the agency as ordered by US health secretary Robert F Kennedy, Jr (RFK Jr), an action some observers believed would cause medical device approval bottlenecks and stifle innovation.
A number of observers perceive Elsa to be a hastily rolled out solution primarily designed to paper over staff cuts at the FDA. In any case, regulatory submissions at the FDA are undoubtedly changing. But Elsa is simply the latest tool in a line of trends at the FDA to streamline efficiencies, following the agency鈥檚 introduction of the electronic Submission Template and Resource (eSTAR) programme, an interactive PDF form that guides applicants through the process of preparing medical device submissions.
genAI: a replacement to human capital?
While initial reactions to Elsa have been mixed, with some observers within the FDA stating that the AI tool 鈥渃onfidently hallucinates鈥� and has 鈥渟evere limitations鈥� since it was trained on , Dr Acacia Parks, a strategic adviser to contract research organisation (CRO) Lindus Health, points out that the majority of reviewers at the FDA were not part of the mass layoffs. Therefore, there is still very much a human-in-the-loop element when AI is reviewing device regulatory submissions.
鈥淯ltimately, a reviewer is the one who’s got to sign off on it,鈥� says Parks.
US Tariffs are shifting - will you react or anticipate?
Don鈥檛 let policy changes catch you off guard. Stay proactive with real-time data and expert analysis.
By GlobalData鈥淚 don’t see a reviewer looking over any letter that a combination of what their review templates and AI have generated and being satisfied if, for example, there鈥檚 no validation data where they were expecting to see it.
鈥淲here we may see gaps, due to layoffs, is in these 鈥榩hone a friend鈥� consult people who aren’t on the review team, but who reviewers turn to when something outside of their purview arises when they are looking at a submission 鈥� such a step may no longer be possible because that expert may or may not be there any longer.鈥�
Dr William Soliman, CEO at the Accreditation Council for Medical Affairs (ACMA), states that while AI at the FDA may streamline the development of content and the management and organisation of content to make submissions more streamlined, there will need to be 鈥渓ots of quality checks鈥� along the way.
Soliman says: 鈥淲e have seen in other fields, like law, for example, where attorneys have lost their privilege to practice because they used AI to prepare legal briefs without checking the accuracy of the information. In some cases, the AI was making up reference cases outright.”
The overuse of AI also means there is a risk in that it could miss critical information points that humans would identify, Soliman says, giving rise to the possibility that less rigorously assessed medical devices could make it on to the market.
Soliman likens the situation to the 鈥渄ebacle鈥� seen at United Healthcare, wherein the healthcare insurance company used AI that automatically denied claims from sick elderly customers, ultimately leading to a
Soliman says: 鈥淭his is a definite risk, especially given that there is currently no regulation on this AI.鈥�
While there is a greater need for streamlined, more concise applications to the FDA, Parks does not view this as a factor likely to result in less rigour in the review process or exploitation and 鈥榗orner-cutting鈥� by those filing submissions.
Parks says: 鈥淚f FDA reviewers feel like they don’t have what they need, they’re going to send a regulatory submission back or reject it, and it鈥檚 more a case that this consequence will be felt on the sponsor side. If a reviewer can’t find what they’re looking for, they’re just going to conclude that it’s not there.
鈥淢y understanding from talking to people on the FDA side informally is that they use internal templates, and if you check all the boxes perfectly, you don’t need 400 pages of documentation; you could do it in 20, in a much shorter manner that’s not worse, just more streamlined.
鈥淚t’s ultimately discounting extraneous information and only including that which is most necessary.鈥�
The Elsa effect
Before eSTAR, there were a lot of stylistic differences in terms of how people decided what to put in a meeting or submission package, says Parks, and how they chose to go about it. With Elsa, sponsors will again have to think more carefully about how they file their submissions to the agency moving forward.
Parks says: 鈥淭he increase in automative tools and AI, along with growing staff shortages at the FDA, are all factors that contribute to the same overall trend, which is that if you are thoughtful and strategic about what you put in your package, you will face less resistance.
鈥淭hat has become truer this year, but it was already becoming truer with the use of eSTAR, where if you put the right piece of information in the wrong part of your submission, and it didn鈥檛 end up in the template, requests for a resubmission would be made.鈥�
While in the past, when describing how, for example, a quality system worked, Parks says that a consultant may have chosen to export everything and send all this detail in a long file to the FDA, just to be safe, but due to the automative developments at the FDA, a more thoughtful approach is now more prescient than ever.
Parks adds: 鈥淧eople who are taking the kitchen sink type approach are quickly going to learn that they need to tighten up their process and make it more streamlined and strategic.鈥�
Far from giving rise to a more lackadaisical approach to device submissions, AI鈥檚 introduction at the FDA appears to have been made with efficiency streamlining top of mind. While on a broad level, there have been deep personnel cuts at the FDA in recent months, review teams have been shielded from many of the recent layoffs, meaning there has not been an appreciable impact on the device review process.
While there may be fewer in-house experts at the FDA, and concerns around over-reliance on AI, it appears that AI is being construed as a tool to augment the review process instead of a tool being applied to take the human out of the loop.
Only time will tell whether Elsa鈥檚 role at the agency will prove effective. As more medical device companies and their sponsors go through the process, and with the FDA acknowledging that Elsa will evolve, so too will how agency reviewers adapt and apply it.
