AI Tool Could Cut Medical Device Recalls by One-Third, Study Finds

The path to market for most medical devices in America is not what patients assume it to be. Rather than undergoing rigorous safety testing from scratch, the vast majority of new devices are approved because they resemble something already on the market—a predecessor deemed safe enough to use. The FDA calls this the 510(k) pathway, and it rests on a simple logic: if Device B is substantially equivalent to Device A, and Device A is safe, then Device B should be safe too. The problem is that similarity and safety are not the same thing. Devices cleared through this mechanism still get recalled. Some prove dangerous in ways their predecessors were not. And the FDA, working with finite resources, reviews each submission the same way regardless of how much actual risk it poses.

A study published in Management Science suggests a different approach: let algorithms do the sorting work, and let human experts focus on the cases that matter most. Researchers from Indiana University, Harvard Kennedy School, and Emerging Health Consulting built a machine-learning system designed to triage FDA submissions. The tool estimates which devices are genuinely low-risk and could be cleared with minimal oversight, which ones are risky enough to reject quickly, and which ones warrant the kind of careful human scrutiny that only an experienced reviewer can provide. The results, if the model holds in practice, are striking: a 40.5% reduction in the time the FDA spends reviewing submissions, a 32.9% improvement in the recall rate—bringing it down from the current 10.3%—and an estimated $1.7 billion in annual savings from fewer device replacements and recalls.

The current system treats all submissions as equivalent problems. An experienced FDA reviewer spends the same amount of time on a straightforward case as on a genuinely novel or complex one. That is inefficient. It also means that the agency's limited pool of expert reviewers gets stretched thin, potentially missing red flags in the cases that actually need their attention. The machine-learning approach inverts this logic. It uses historical data on which devices got recalled and which did not, along with characteristics of the submissions themselves, to predict which new submissions are likely to end up in trouble. Devices that the algorithm flags as safe bets could be cleared quickly or even automatically. Devices that look risky could be rejected without extensive review. And the devices in the middle—the ones where the risk is genuinely hard to assess—would get routed to human experts who could spend real time on them.

This is not about replacing human judgment. It is about directing it where it is most needed. The researchers built the system to work alongside FDA reviewers, not instead of them. The algorithm makes a recommendation; humans make the final call. But by handling the triage work, the system frees up expert time for the submissions that actually require expertise. The study found that this hybrid approach could catch more unsafe devices than the current system while simultaneously reducing the burden on reviewers. It is a rare outcome in policy work: doing more with less, but only because the less is being spent more wisely.

The question now is whether the FDA will adopt it. The agency has shown interest in using data and algorithms to improve its processes, but deploying a new system at scale requires not just technical validation but institutional buy-in, staff retraining, and the kind of careful implementation that takes time. The study provides the evidence that the approach works in theory. Whether it works in practice—whether the algorithm's predictions hold up when applied to real submissions, whether reviewers trust the system enough to act on its recommendations, whether the promised savings actually materialize—remains to be seen. But for a system that has relied on the same basic logic for decades, the possibility of a smarter way to allocate resources is worth serious consideration.