More than 40% of artificial intelligence (AI) tools cleared or approved by the Food and Drug Administration (FDA) from 1995 to 2022 lack clinical evidence, as highlighted in a recent Nature Medicine article. Researchers from the University of North Carolina School of Medicine studied 521 FDA-authorized AI medical devices, finding that 226 devices (~43%) lacked clinical validation data. These findings underscore concerns regarding the reliance on less rigorous regulatory pathways and emphasize the importance of conducting prospective studies and publishing clinical validation data to ensure the safety and effectiveness of AI technologies in healthcare.
The FDA typically clears most AI medical devices through its 510(k) process. This process requires manufacturers to demonstrate that their new products are substantially equivalent to a previously approved device, termed a predicate device. This pathway focuses less on individual validation data, assuming safety and effectiveness based on similarity to existing technologies. Sammy Chouffani El Fassi, UNC researcher and study’s first author, stated in an interview that although many of these technologies legally did not require independent validation data, lacking such evidence could potentially slow their adoption.
Chouffani El Fassi, an M.D. candidate at UNC, further explained the importance of validating AI medical devices in clinical settings. Small modifications to an algorithm, he noted, could significantly alter a device’s impact after deployment, hence testing in a real-world environment is crucial. The research team categorized AI devices based on the presence of clinical validation, and whether they were tested retrospectively or prospectively. Out of the 292 devices with validation data, nearly half underwent retrospective testing (usually before actual patient implementation), while the other half were validated prospectively with real-time patient data during trials.
Prospective studies, involving only 22 randomized controlled trials among the devices reviewed, are especially valuable as they provide deeper insights into how a device performs under current and realistic conditions. Chouffani El Fassi shared experiences from a Duke University team’s project validating an algorithm designed to detect cardiac decompensation from electronic health records. Initial retrospective studies were complemented by a prospective study where cardiologists used the algorithm and provided feedback on its accuracy and usability in real-time scenarios. This approach highlighted necessary improvements, such as enhancing the user interface for better efficiency and user experience.
Prospective testing also helps identify potential confounding variables that retrospective data might not reveal. For instance, Chouffani El Fassi discussed how the onset of the COVID-19 pandemic altered chest X-ray appearances, which would not have been captured if the device was only tested using pre-pandemic data. Furthermore, prospective validation doesn’t have to be complex; even simple methods such as clinicians rating the usefulness of an AI-optimized ultrasound probe in practice can yield valuable insights, exposing real-life issues such as poor user interfaces that might not be apparent in controlled test environments.
Chouffani El Fassi advocated for more straightforward and practical approaches in conducting prospective studies, emphasizing that real-world testing can reveal crucial information about an AI device’s performance and user interaction. This approach is seen as the most beneficial form of data for validating medical AI tools, as it captures the nuances of everyday clinical encounters and directly addresses the practical challenges faced by healthcare providers.
Overall, the findings from the UNC researchers call for an increased focus on comprehensive and rigorously-collected clinical validation data for AI medical devices. By prioritizing prospective studies and making validation processes more rigorous and transparent, healthcare providers can better assess the true value and reliability of AI technologies, ultimately ensuring safer and more effective patient care.
#FDAauthorized #devices #lack #validation #data
