Publicly available data from the Food and Drug Administration pertaining to regulated artificial intelligence and machine learning (ML) algorithms are in short supply, creating challenges for radiologists and other doctors to utilize these technologies clinically.
“Although FDA maintains a searchable platform for its regulated devices including the regulated AI/ML algorithms, it does not offer an accessible list of regulated software,” corresponding author, Shadi Ebrahimian, MD, with the Department of Radiology at Massachusetts General Hospital and Harvard Medical School, and co-authors explained. “Therefore, there is confusion regarding the type and nature of the regulated algorithms.”
In 2010, there were around 600 publications that referenced AI/ML. That number has since climbed to more than 12,000. In recent years, the number of AI/ML applications regulated by the FDA has also increased rapidly, though very few of these specific applications have been referenced in peer-reviewed research.
To develop a better understanding of the strengths and trends of the software that is regulated by the FDA, doctors analyzed the FDA database of 510(k) clearance documents and corresponding product websites. They then recorded the details of the application type, body area of application, algorithm performance and validation studies.
The research, published in Academic Radiology, revealed that of the software regulated by the FDA, 93% (118/127) is either an AI or ML algorithm. Out of those, 17 did not share validation claims or data. Only 9 of the 118 had validation datasets of over 1,000 patients, with the majority having less than 500 patients. Most lacked any patient demographic information.
“Several of these algorithms (14%) had no mention on any details on validation studies or patient datasets used for training or testing to draw meaningful conclusions on their generalizability or robustness,” the doctors revealed.
They found that the most common function of AI/ML applications regulated by the FDA was quantification, followed by detection. The applications were geared towards CT or MRI and typically focused on one specific body region, with the most targeted anatomy being the brain, breasts and lungs.
“The main implication of our study is that FDA summaries lack sufficient information on validation of AI/ML algorithms,” the doctors explained. "It suggests that companies with FDA-regulated AI algorithms should provide additional details on their validation data, perhaps including this on the ACR website.”
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She joined Innovate Healthcare in 2021 and has since put her unique expertise to use in her editorial role with Health Imaging.