Artificial intelligence-enabled medical devices cleared by the FDA are more likely to be recalled if they come from publicly traded companies, according to a new study.
Researchers from Johns Hopkins University and Yale examined 950 AI-enabled medical devices approved and cleared by the FDA as of late 2024. They found that many entered the market with little or no clinical validation. The findings were published in JAMA Health Forum.
Publicly traded companies accounted for 53.2% of the devices reviewed but were linked to more than 90% of recall events. Public company status was associated with a 5.9 times higher chance of recall. Among recalled devices, 40% from private firms lacked clinical validations compared with 77.7% from established public firms and 96.9% from smaller public firms.
“The lopsided nature of these recalls should give every advocate for medical AI pause. Publicly traded companies, the big fish in this still-small pond, built just over half the devices but were responsible for nearly all the recalled units,” said Tinglong Dai, corresponding author and professor at the Johns Hopkins Carey Business School.
The researchers pointed to the FDA’s 510(k) clearance pathway, which does not require prospective human testing. Nearly half of all AI device recalls in the study occurred within the first year after clearance, double the rate reported for all 510(k) devices. The authors said this suggests the pathway may overlook early performance failures of AI technologies.
“We were stunned to find that nearly half of all AI device recalls happened in the very first year after approval,” Dai said. “And the even bigger surprise? The recalls were heavily concentrated among tools with no reported clinical validation whatsoever. We just thought, ‘wow — if AI hasn’t been tested on people, then people become the test.’”
Medical AI raises broader concerns
To mitigate risks, the authors suggested requiring human testing or clinical trials before clearance, or approvals that expire unless companies provide evidence of real-world performance. They also recommended stronger post-market surveillance and incentives for firms to conduct ongoing studies.
“If these tools are going to scale, we’ve got to test them on real people,” Dai said. “Either require proper human studies up front or give approvals that expire unless the evidence shows they actually work.”
The findings add to growing concerns about unintended consequences from AI in medicine. Earlier this month, a study in The Lancet Gastroenterology & Hepatology found that doctors detected fewer polyps when performing colonoscopies without AI assistance after becoming accustomed to working with AI support. That research raised questions about “deskilling,” or the loss of fundamental clinical skills when clinicians over-rely on AI.
A separate study published in npj Digital Medicine in October 2024 examined 692 FDA-cleared AI/ML devices approved from 1995 to 2023 and found widespread gaps in transparency. Fewer than 4% of approvals reported race or ethnicity data, and more than 80% did not report patient ages. Fewer than 10% included post-market surveillance studies. The authors warned that incomplete demographic reporting could worsen algorithmic bias and health disparities.
