Study: Data lacking for AI-enabled medical devices cleared by FDA

Clinical performance data for medical devices with artificial intelligence (AI) features is often unavailable, making it difficult to assess real-world outcomes and potentially jeopardizing safety, according to a new study of products cleared by the US Food and Drug Administration (FDA).

The study looked at 903 products, the majority of which were used in radiology applications (76.6%). Clinical performance studies were available for a little more than half (505, or 55.9%) at the time of FDA clearance, according to a JAMA Network Open report published April 30.

Consequently, it’s hard to predict performance issues over time and with wider use, George Siontis, a cardiologist University of Bern in Switzerland, and colleagues wrote.

“An AI device developed and validated in one country or region might not perform equally well in another with a different patient population, potentially impacting the generalizability of AI models on a global level,” the authors advised. “Therefore, the clinical evaluation and validation of AI-enabled medical devices is crucial and remains challenging.”

Mining public data

In light of the rise of AI across industry sectors and awareness of shortcomings in available data for medtech, the study was designed to evaluate the quality of publicly available supporting the 510(k) clearances and de novo/ premarket approvals of AI-enabled devices used for healthcare indications. Researchers assessed all 903 products on the FDA’s list of AI and machine learning (ML) devices published from inception through 31 August 2024, with 877 products cleared via the agency’s 510(k) pathway and 22 de novo approvals. Researchers noted that 664 of the 903 were purely software products (73.5%) and the rest were physical devices with AI software.

“The number of AI-enabled devices approved by the FDA has grown exponentially in recent years,” Siontis et al. noted.

By medical specialty, the devices and software products included radiology (692, 76.6%), cardiology (91, 10.1%), and neurology (29, 3.2%).

Researchers found that in almost one-quarter of submissions to FDA, sponsors indicated no clinical performance study had been conducted, and in almost 20% it was not specified whether a clinical performance study had been completed.

In the 55.9% that completed clinical performance studies, retrospective was the most common trial design (38.2%). But the type of study design and other key information were often not specified.

“Among the clinical performance studies, information on sex subgroups was available in less than one-third (145 studies [28.7%]), and 117 studies (23.2%) provided information on age-related subgroups,” the authors noted.

“Ensuring that AI models are trained on comprehensive and representative datasets is essential to avoid biases and ensure reliable outcomes in real-world applications,” the authors advised. “AI models are highly dependent on the data on which they are trained.

Setting stage for recalls?

Furthermore, important metrics like sensitivity, specificity and area under the curve (AUC) were typically not available, they added.

Out of the total 903 products cleared through August 2024, 43 (4.8%) were recalled and the average time between approval and recall was 1.2 years. The authors warned that effectiveness and safety could be compromised when devices are used widely.

Siontis et al. wrote that the reliance on the 510(k) pathway “raises significant safety concerns, particularly for high-risk (class I) medical devices.” Despite the risks, all implantable AI-enabled medical devices were 510(k)-cleared, they noted.

“Notably, 19% of class I recalls across all medical devices involved implantable devices, potentially highlighting that substantial equivalence may be an inadequate marker of safety and efficacy,” they wrote.

Going forward, to remedy the shortcomings of publicly available data, industry, academia and regulatory bodies should work together to improve the landscape, the authors suggested. (RELATED AdvaMed: AI medtech companies facing more risk with uncertainties at FDA, Regulatory Focus 12 April 2025)

AI-enabled medical devices promise to transform healthcare in the future by improving efficiency, personalization and access, the authors acknowledged.

“Achieving this vision will require sustained collaboration among technology companies, clinicians and regulators to ensure the safe and ethnical application of AI in clinical practice,” the researchers wrote. “The medical community should play a central role in this process, which requires them to receive adequate training to effectively monitor the performance of AI-enabled medical devices in real-world settings on an ongoing basis.”

Study: Data lacking for AI-enabled medical devices cleared by FDA

Related topics

Recommended content

Welcome to the new RAPS Digital Experience