Study Says Machine Learning Could Help FDA Target Riskier 510(k) Submissions
Researchers modeled a human-plus-algorithm review policy on more than 31,000 submissions and estimated a 40.5% cut in reviewer workload alongside fewer later recalls.
.jpg){kind=link}
Key points
- A Management Science study proposes a machine-learning model that sorts 510(k) submissions into likely clear, likely reject and refer-to-human-review groups.
- In the authors' conservative evaluation, the policy improved the recall rate by 32.9% and cut a normalized measure of FDA workload by 40.5%.
- The model leans on the age and recall history of predicate devices; the authors estimate about $1.7 billion in annual savings from avoided replacements.
A machine-learning system that estimates a device's chance of being recalled could help the FDA decide which 510(k) submissions need close human review and which do not, according to a study in the journal Management Science that its publisher, INFORMS, publicized on September 21. The authors estimate their approach would cut the agency's review workload by 40.5% while improving the recall rate by 32.9%.
The 510(k) pathway lets a manufacturer bring a device to market by showing it is substantially equivalent to a legally marketed "predicate" device. The authors argue that the ambiguity in that test has been associated with high recall among many devices cleared this way, according to the paper's abstract.
How the model works
The researchers, from Indiana University, Harvard Kennedy School and Emerging Health Consulting, first built machine-learning methods to estimate each submission's risk of recall using information available when it is filed. They then designed a data-driven policy that recommends one of three outcomes: acceptance, rejection, or deferral to FDA experts for in-depth evaluation.
The study draws on a dataset of more than 31,000 submissions that the authors assembled from FDA and Centers for Medicare and Medicaid Services data. A key input is the history of the predicate devices that a submission relies on, in particular how old they are and whether they have been recalled, INFORMS said.
Compared with current FDA practice, which the authors put at a recall rate of 10.3%, a conservative evaluation of their policy showed a 32.9% improvement in the recall rate and a 40.5% reduction in a normalized measure of workload, according to the abstract. The authors also estimate annual savings of about $1.7 billion for the health care system from avoided device replacement costs, which they describe as equal to 1.1% of annual US medical device spending. Those figures are modeled estimates from historical data, not results from a policy the FDA has used.
Human judgment kept in the loop
"AI is most useful in this setting when it works alongside human judgment, not when it tries to replace it," co-author Soroush Saghafian said in the INFORMS release.
"The goal isn't to replace FDA reviewers—it's to give them better information about where their time and expertise can have the greatest impact," said lead author Mohammad Zhalechian.
INFORMS tied the work to a 2023 FDA draft guidance on best practices for choosing predicate devices, which proposed weighing a predicate's safety track record. The study's model tracks predicate recall history and age, the release said.
The paper does not describe any FDA plan to adopt such a tool, and the agency was not quoted in the release.
Sources
- AI Could Cut Medical Device Recalls by Nearly a Third, Study Finds, INFORMS (publisher of Management Science), September 21, 2026
- Harmonizing Safety and Speed: A Human-Algorithm Approach to Enhance the FDA's Medical Device Clearance Policy (arXiv v3), arXiv (Zhalechian and co-authors), January 29, 2026
- Medical Device Recalls, U.S. Food and Drug Administration, September 23, 2026