Researchers Use AI to Better Predict Medical Device Recalls
Thursday, May 28, 2026
By Rose Semenov
A deep learning predictive model developed in part by Minnesota Carlson researchers could potentially save lives and millions of dollars by improving the prediction accuracy of medical device recalls.
The research, published in Information Systems Research, was co-authored by Yi Zhu, ’24 PhD, Information & Decision Sciences Associate Professor Soumya Sen, Finance Professor Pinar Karaca-Mandic and Alexander Everhart of Washington University. Zhu led the study as a Carlson PhD student and is now an assistant professor at the University of Texas at Dallas.
The vast majority of medical devices gain clearance through the U.S. Food and Drug Administration’s 510(k) pathway. This pathway allows manufacturers to prove that their device’s technology is “substantially equivalent” to one or more predicates — existing devices that are already available on the market. While the 510(k) pathway allows for variations of devices to enter the market more quickly, studies have shown these devices are more likely to be recalled, which can have negative impacts on patients and manufacturers.
The researchers developed a model, called DeepPredicate, which novelly incorporated predicate network characteristics — associations between a device and its predicates over time — in addition to other device factors, to enhance its recall predictions. To test DeepPredicate’s effectiveness, they used data from 45,398 medical devices cleared between 2003 and 2020. They focused on predicting devices’ first Class I or Class II recall, the most serious levels, after clearance. They determined:
- DeepPredicate improved the prediction recall score by 6.56 percentage points compared to traditional models.
- The researchers estimate that improved score would save the industry roughly $328 million annually by setting proactive responses in motion for correctly predicted recalls.
- The DeepPredicate model can be adapted to examine other complicated networks, such as traffic or social networks.
“This model would help healthcare providers, manufacturers and regulators better assess the recall risk of medical devices, which could ultimately improve patient care,” said Karaca-Mandic, the C. Arthur Williams Jr. Professor in Healthcare Risk Management. “These AI advancements can help power social good and mitigate risk within the industry.”
Using the DeepPredicate model, the researchers created the online tool Medevisor, a prototype that evaluates medical device safety based on recall data. The tool is currently available for research purposes only and does not provide any clinical recommendations. The researchers note that the DeepPredicate model is meant to augment, not replace human expertise and that its predictions must be evaluated and confirmed.
“The human component remains an important part of the evaluation process,” said Sen, who holds the 3M Excellence in Business Analytics Fellowship. “When used with the proper oversight, a model like this can demystify the complex patterns within the 510(k) device network, lead to earlier recall identifications and enhance response.”