Most medical devices reach the United States market not through rigorous safety trials but by being judged substantially equivalent to an already cleared product under the FDA’s 510(k) pathway. Similar, as the study’s authors put it, is not a synonym for safe.
A new paper in Management Science models what happens when a machine-learning system estimates recall risk across thousands of submissions and their predicate histories, then pairs that estimate with a decision policy. The tool sorts filings into three piles: likely safe, risky enough to reject after brief scrutiny, and uncertain enough to send to human experts.
The reported results: a 40.5% reduction in review workload, a 32.9% relative improvement in the recall rate – the current rate sits at 10.3% – and up to roughly $1.7B in annual health care savings from fewer device replacements.
Researchers from Indiana University, Harvard Kennedy School and Emerging Health Consulting built the model. Co-author Soroush Saghafian framed the design choice plainly: the tool is meant to work alongside human judgment, not replace it.
The proposal lands as regulators face a growing queue of submissions, including AI-enabled devices, and as critics argue the 510(k) route lets weak evidence through. Automating the easy calls would free reviewer time for the hard ones.
Whether the FDA adopts anything resembling this model is a separate question. The study’s contribution is showing that triage, rather than blanket scrutiny, may be the cheaper path to catching dangerous devices earlier.
