A peer-reviewed study found that 76.5% of 1,430 FDA authorization records for artificial intelligence and machine learning-enabled medical devices were reviewed by the agency’s Radiology panel. The analysis, led by GigHz founder Pouyan Golshani, MD, examined records dated from September 1995 through December 2025.
Radiology accounted for 1,094 records. The Radiology, Cardiovascular and Neurology panels together accounted for 90.6% of the authorizations, according to the study published in Cureus. The results describe authorization patterns; they do not show what caused the concentration.
Golshani, an interventional radiologist and the study’s lead author, pointed to radiology’s existing digital infrastructure as one possible explanation. He said medical images, common file standards and systems for moving scans to readers give developers a place to integrate AI. He cautioned that this does not mean radiology is easy or that radiologists are close to being replaced.
Authorization Growth and Developer Landscape
The authors reported 331 authorizations in 2025. Annual authorizations averaged 1.8 from 1995 through 2014, compared with 264 per year from 2023 through 2025. The study counted nine records under Pathology, six under Microbiology and four under Obstetrics and Gynecology across the full period.
The analysis also found that many companies had only one listed AI device. Of 740 companies, 502, or 67.8%, had a single authorized device. Thirteen companies, or 1.8%, accounted for 247 devices, representing 17.3% of the total.
No authorizations appeared under a psychiatry or behavioral health review panel. The authors note that FDA review-panel categories do not map directly to every specialty or care setting. A low count in a category does not establish that a specialty lacks AI tools.
From Authorization to Clinical Use
Golshani said other fields, including internal medicine, have guideline-based decisions where better support could help. But relevant information may be spread across notes, laboratory results, medications and prior visits, he said. His comments describe a development challenge, not a finding tested by the study.
He argued that developers and health systems should define a specific clinical task, make necessary data available and test a tool in the workflow where it will be used. A documentation tool and a system that recommends treatment require different evidence and safeguards, he said.
Golshani said both fear of replacement and fear of missing out can lead to poor decisions. He called for testing what a tool improves, where it fails and who is responsible when it does. Those are his views; the analysis did not assess device performance, clinical adoption or patient outcomes.
The study measures FDA authorization records, not patient benefit or physician replacement. The FDA’s AI-enabled device list is not comprehensive and does not capture all healthcare AI, including software functions outside device regulation. The authors’ findings therefore concern the records they analyzed, not the full range of AI use in healthcare.
The study was led by Golshani and Mary S. Joseph and published July 13, 2026. Its reported authorization period ends in December 2025. GigHz describes itself as a physician-founded software and research company developing clinical decision support, radiology reporting and practice intelligence tools. The study did not evaluate or validate GigHz products.
The source report identifies the study’s journal reference as Golshani P, Joseph MS, “Three Decades of Food and Drug Administration Authorizations of Artificial Intelligence/Machine Learning-Enabled Medical Devices: Persistent Specialty Concentration and the Care-Delivery Gap (1995–2025),” Cureus. The study’s implications remain bounded by its focus on authorization records and the FDA list’s stated limits.
