TLDR

  • Only 12.2% of 3,967 clinical prediction studies shared their analytical code.

  • Code sharing rose from 6.3% in 2015 to 15.8% in 2025.

  • Just 35.3% of shared repositories included a license, and 3.9% included tests.

  • Some journals with code-sharing policies still had low sharing rates.

  • The findings are shaping TRIPOD-Code, a new guideline for sharing research code.


Most research code shared alongside medical prediction studies is missing key details others would need to run it, a new review has found. Only about one in three code collections included a license, and fewer than 4% included tests.

The review, published in Nature Medicine, looked at 3,967 studies that built, updated or tested clinical prediction models. These models estimate patient outcomes, such as the risk of developing a disease, and help guide care decisions.

An international team led by researchers at MIT and Dana-Farber Cancer Institute carried out the work. They used large language models to read the papers, find statements about shared code and pull out links to online code collections, known as repositories.

Few Studies Share Code at All

Only 482 of the studies, or 12.2%, shared their code. Most did so through a repository link, while 75 placed code in an appendix or supplementary material.

Sharing has grown slowly over time. The rate rose from 6.3% in 2015 to 14.5% in 2024 and 15.8% in 2025.

Studies citing TRIPOD+AI, a 2024 reporting checklist that encourages code sharing, shared code far more often. In 2025, 29.7% of those studies shared code, compared with 11.8% of studies citing the original TRIPOD checklist.

Rates also differed widely by journal. Nature Communications led at 71.4%, while 17 of 71 journals with more than 10 studies in the review had no code sharing at all.

Some journals with low sharing rates still had written code-sharing policies. The authors said this suggests a policy on its own may not be enough.

By country, Finland had the highest rate at 33.3%, followed by Belgium at 26.7% and Israel at 23.1%. China produced the most studies, accounting for 32.4% of the total.

Shared Code Often Falls Short

The team examined 380 repositories in detail, checking each against 14 features linked to reproducibility. GitHub hosted 83.4% of them.

Most repositories, 80.5%, had a README file. But only 52.1% had one that explained the code’s purpose and expected results.

Software requirements were listed in 37.6% of repositories, and only 21.6% stated exact versions. Licensing details appeared in 35.3%, while tests were present in just 3.9%.

Thomas Sounack, the study’s first author, said good code sharing means “giving enough information to understand what the code does and be able to run it to reproduce the results reported in the paper.” He said some repositories were well organized, while others gave almost no guidance.

Python overtook R as the most common programming language in 2023. By 2025, it made up more than half of all languages used.

Two human reviewers checked the AI’s work on 500 papers. The model reached 92.3% accuracy on finding repository links, and the authors said results should be read as overall trends rather than exact counts.

The findings are informing TRIPOD-Code, a new reporting guideline for sharing analytical code. “We want to give researchers practical guidance on how to share code, and give journals clearer expectations to communicate to authors,” said senior author Tom Pollard.

An expert consensus process will now decide which code-sharing elements should be considered essential, recommended or context dependent.