Overview
- A JAMA research letter and AAIC presentation show that FDA summaries for 14 of 24 AI dementia devices omit training-set details while 22 lack validation-set information.
- Public records for 23 systems offer no race or ethnicity data and most also fail to report age, sex or disease status among participants.
- Peer-reviewed publications fill some gaps but report training data for only five tools and validation data for just ten, leaving the majority unverified.
- More than 90% of the devices were cleared through the FDA’s 510(k) process and reviewed mainly by radiology panels under voluntary demographic guidance.
- Researchers warn that these transparency gaps hinder evaluation of algorithmic bias and equitable performance across diverse patient groups.