BIOBRIEFPAPER
TRAM AI (Transparency Reporting of Medical Imaging Datasets for AI) gives medical imaging AI a 14 item audit; on 10 pilot datasets, basic de identification hit 90%, long term archive validation just 10%.
Medical AI papers routinely claim their datasets are "publicly available." A peer-reviewed audit of 10 such datasets found that label is doing a lot of work: basic de-identification passed 90% of the time, while long-term archive validation passed just 10%.
The instrument is TRAM-AI (Transparency Reporting of Medical Imaging Datasets for AI), a 14-item consensus checklist published in Radiology: Artificial Intelligence (RSNA, 2026; PMID 42905015). The items cover four domains: dataset access and persistence, technical interoperability, metadata and reproducibility, and ethical, legal, and maintenance. An international panel of 33 built the list across two modified Delphi rounds, with consensus defined as at least 75% agreement on both criticality and feasibility.
Compliance on the 10 pilot datasets varied sharply. Trusted-repository and subject-level-mapping items hit 90%; partition keys reached 30%. Archive validation passed 10% of the time. It is the check that decides whether outside labs can reproduce a study years later.
The authors position TRAM-AI as a complement to existing methodological reporting standards, not a replacement, and call for broader validation. A four-person advisory group from Latin America and sub-Saharan Africa separately endorsed the checklist, extending the review base beyond the 33-expert panel.