AI Enters The 510(k) Submission World: What Device Developers Need To Know

By Jim Kasic, Boulder iQ

AI has reshaped lives and quietly integrated itself into all types of systems, and medical devices are no exception. Beyond the surge in development of AI-enabled devices over the past few years, AI is now impacting the FDA’s 510(k) submission process for medical device developers.

The Digital Move

In 2023, the FDA made the electronic Submission Template And Resource (eSTAR) mandatory for most 510(k) submissions. Although eSTAR is not an AI tool, the move was the first definitive leap toward the agency’s goal of making the review process more digital and more streamlined.

In this format, eSTAR allowed all information provided in submissions to be in an online, organized format, which reduced the need for manual intervention. Before the introduction of eSTAR, those involved in 510(k) submissions would spend valuable time responding to the same questions from multiple reviewers who were not able to share different files or sections of the submissions.

With eSTAR as a framework, the FDA now is moving toward agency-wide AI integration, with the goal of further increasing the efficiency and speed of its review process, including 510(k) applications. The FDA has emphasized that the AI tools are meant to support, not replace, human reviewers by reducing repetitive tasks that can slow down the review process and allowing a shift of focus toward substantive scientific review.

In May of this year, the FDA implemented an AI-assisted scientific review pilot within its Center for Drug Evaluation and Research (CDER). The program was intended to accelerate reviews by having AI draft summaries and flag missing information. In June, the next step included scaling the use of the technology across all FDA centers (drugs, devices, biologics, and others). While all centers are using a shared, secure, AI platform to support internal operations and review workflows, the system will continue to be tailored to fit the needs of each center as it continues to evolve.

FDA’s AI Tool: Elsa

The FDA’s implementation of AI tools marks the next step in the agency’s digital transformation efforts, which is indeed a step and a start, albeit far from a finished system. The agency is currently employing two key tools in its AI implementation:

• Generative AI copilots. These tools improve efficiency by helping reviewers with tasks like technical screening, document summarizing, and organizing large document packages. Elsa (“Electronic Language System Assistant”), the name for the FDA’s designated AI copilot, is designated as an “efficiency-enhancement” tool, helping agency employees with tasks like reading and summarizing internal documents. It was built on a secure platform and operates within Amazon Web Service (AWS)’s secure GovCloud environment. According to the FDA, it does not train on industry data or any submissions.
• Structured data analysis. The FDA is encouraging manufacturers to use structured formats and clean, consistent data in their submissions. This allows the agency's AI tools to more efficiently parse and analyze the content of a submission.

Impact On 510(k) Submissions

AI processes have not been in use for a long enough period of time to accurately ascertain the full impact on the 510(k) submission process. While any new system can be expected to experience growing pains, we have seen some early concerns from staff on Elsa’s performance. And a limited pilot period and little information about training or testing processes for Elsa make it impossible to fully understand its real-world performance yet.

That said, we can take an optimistic view and see that medical device developers stand to benefit from the FDA’s use of AI tools in a number of ways. By freeing reviewers to focus on the most complex, highest-risk issues, AI’s handling of routine tasks has the potential to reduce timelines for 510(k) filings and cut down on the number of resubmission cycles needed for clearance. 

AI tools can:

• Automate completeness checks and spot gaps early on
• Streamline the review of technical documentation
• Identify potential issues earlier. This could lead to a reduction in the number of “Additional Information" requests, which historically have placed submissions on hold and delayed clearance. 
• Speed up technical screening of large-evidence packages

The bottom line is that AI could change how quickly applications move through the FDA, ultimately allowing developers to bring their devices to market — to providers and, ultimately, patients — faster.

What To Watch For

As AI becomes increasingly integrated into the 510(k) process, those involved in device development have (and should have) questions.

• Will 510(k) sponsors be notified when AI has a role in the review of their submission?
• If a developer experiences delays, or does not receive clearance on a device, and AI was involved, what kind of visibility will that developer have into the process?

We know, too, that just as AI learns and builds on good data, it can do the same with bad, or irrelevant, data. One example from my memory is a situation where AI analyzed clinical data on performance of a device – and came up with a key finding that all the patients in the data set lived on streets with names of trees (e.g., Maple, Ash, Elm). While it may have been true, the finding was clearly not relevant. What systems will be set up to allow for needed dialogue to resolve inconsistencies or discrepancies?

What You Can And Should Do In Your 510(k) Submissions

To leverage and benefit from the FDA’s use of AI in 510(k) submissions, developers can prepare submissions in a way that’s most conducive to the AI system. In other words, submissions should be as “AI-ready” as possible.

• Use structured formats and clean data. To be accurate and efficient, AI needs consistent formatting and labeling. Device developers with experience submitting in eSTAR will be a step ahead in providing data that AI tools can efficiently parse and analyze.
• Organize by topic. It’s human nature to think in chronological terms, but presenting information chronologically is generally not the way to streamline your submission review when it comes to how AI tools work. Instead, structure your submission so that it groups together related materials in a logical format – more like a table of contents than a story or a day-to-day log.
• Watch out for, and eliminate, redundancy and inconsistencies. Any reviewer — human or AI — will stumble when encountering conflicting information across documents in a submission. Establish consistency as a priority in your quality control system. Eliminate possibilities for slowdowns with careful internal review and proofing of submissions.
• Document assumptions. As always, summaries and introductions to modules can identify key points, explain an approach, and help reviewers follow your logic. Strategic documentation is critical, whether a human or AI reviewer is reading a submission.
• Maintain legal counsel involvement. Legal counsel remains essential in assuring that submissions meet FDA standards. They also will be able to anticipate AI-driven feedback and assessments.

Cautious Optimism Moving Forward

To date, much of the AI review process has focused on the pharma industry. As the FDA refines the process, we expect to see tips, guidance, and more specific information for device submissions, particularly for visuals (device diagrams and images) and file formats.

Elsa is the beginning, not the end. It will not be the only AI tool the FDA uses in the realm of 510(k) submissions but rather part of a broader shift across government agencies to integrate AI into operational processes. While Elsa may indeed increase productivity, it is laying groundwork – and comes with unanswered questions and unknown results.

It will be critical for device developers to monitor progress and developments closely as they start to make 510(k) submissions, then ask — and demand — answers.

About The Author:

Jim Kasic is the founder and chairman of Boulder iQ. With more than 30 years of experience in the Class I, II, and III medical device industry, he holds more than 40 U.S. and international patents. His career includes experience with companies ranging from large multinational corporations to startups with a national and international scope. Kasic has served as president and CEO of Sophono, Inc., a multinational manufacturer and distributor of implantable hearing devices, which was acquired by Medtronic. He also was the president of OrthoWin, acquired by Zimmer-BioMed. He received a B.S. in physics and an M.S. in chemical/biological engineering from the University of Colorado and an MBA from the University of Phoenix. He can be reached at jim.kasic@boulderiq.com or on LinkedIn.