Availability is only the first question
A link to research code can answer where the files are. It cannot, by itself, tell another team how to run them. MIT’s 19 November account of an AI-assisted review brings that gap into focus. The underlying Nature Medicine paper appeared in October; the useful question is what its findings mean for readers assessing a published model.
Clinical prediction studies develop or test methods for estimating an outcome from several pieces of information. To understand the reported result, another researcher may need to trace data preparation, model settings and evaluation steps. A repository—the online collection of code and related files—can help expose those choices, but only if it is usable.
What the review counted
Thomas Sounack and colleagues examined 3,967 eligible papers and identified code sharing in 482, or 12.2%. They characterized 380 repositories against predefined features. Their sample was restricted to studies citing TRIPOD or TRIPOD+AI reporting guidance and available through PubMed Central’s open-access collection; it does not represent every medical study.
The workflow used language models to screen papers, find code statements and support repository assessment. MIT reports that two researchers independently checked 500 papers and a separate repository sample. It also notes that automated judgments agreed more closely with people on concrete file characteristics than on the adequacy of documentation.
What a useful code release needs
The paper distinguishes a README file’s presence from instructions that explain purpose and expected output. It also examines software dependencies, licensing, tests and sample data. Checking those features is not the same as rerunning every analysis or validating a model on new patients.
For a reader, a more informative question than whether code is online is whether the release explains how to use it and what should happen when it runs. The authors acknowledge classification errors and limits to generalization. Their review informs the developing TRIPOD-Code guidance; it does not certify individual models as reproducible, accurate or safe.
Sources
- Sounack and colleagues: AI-assisted scoping review of code sharingFull publisher-version paper hosted by the University of Birmingham; DOI 10.1038/s41591-026-04691-1, selection criteria, repository assessments, methods and limits.
- MIT News: How medical studies share code19 November account, human validation, documentation issues and TRIPOD-Code's work toward reporting guidance.
- Dana-Farber: Original dated research releaseCo-author institution’s 9 October research release independently confirms the paper’s earlier publication and named research team.



