Summary
A Nature Medicine scoping review found that 482 of 3,967 studies citing TRIPOD or TRIPOD+AI reported sharing analytical code. Of 380 repositories the researchers could characterize, documentation, dependency details and tests were often missing.
A scoping review published in Nature Medicine found that 482 of 3,967 clinical prediction model studies citing TRIPOD or TRIPOD+AI reported sharing analytical code. The researchers could retrieve and examine 380 repositories; many lacked features that make software easier to understand and rerun, including clear documentation, specified software dependencies and tests.
The study used a large-language-model-assisted process to identify relevant papers, locate code-sharing statements and assess repository features. Its findings provide a baseline for TRIPOD-Code, a reporting guideline extension in development focused on code availability and computational reproducibility.
Code-sharing reports were uncommon
Clinical prediction models combine information from multiple predictors to estimate the probability of a condition or future outcome. Their development and evaluation involve analytical steps such as data processing, model fitting and performance testing. Access to the code can help other researchers examine those steps and reproduce the reported results.
The review covered PubMed-indexed papers citing TRIPOD or its 2024 update, TRIPOD+AI, as of 11 August 2025. After screening and eligibility checks, 3,967 studies were included. The researchers classified 482, or 12.2%, as reporting code sharing. That total included code provided in supplementary material and links to repository providers their retrieval tool could not assess, as well as repositories they could retrieve.
The reported rate increased over time: it was 6.3% in 2015 and 14.5% in 2024. Papers citing TRIPOD+AI had a higher code-sharing rate than those citing TRIPOD alone, at 29.2% and 11.4%, respectively. The association remained after adjustment for publication year. Because this was a comparison between groups of published studies, it shows an association, not that the guideline update caused authors to share more code.
Repositories often lacked practical reuse features
Of 447 repository links identified, 380 repositories could be retrieved and characterized. Most were hosted on GitHub. Although 80.5% had a README file, only 52.1% had one that described the repository’s purpose and expected outputs. The researchers classified 42.4% as modular and structured. Software dependencies were specified in 37.6%, and version constraints were provided in 21.6%. Tests were present in 3.9% of repositories.
These features affect whether another researcher can work out what the software is intended to do and recreate the computational environment needed to run it. Code can be publicly visible while still being difficult to install, interpret or execute. The review therefore distinguishes reporting that code is shared from the practical characteristics of the shared repositories.
The study’s LLM-assisted screening and extraction process was checked against human annotations. On 500 randomly selected articles, it achieved a weighted F1 score of 0.97 for selection criteria and 92.3% accuracy in retrieving repository links. Repository-feature extraction was evaluated against human labels for 35 repositories, with a weighted F1 score of 0.83. These checks support the use of automation at scale while indicating that classification, especially of more subjective features, is not error-free.
The sample was limited to studies citing TRIPOD or TRIPOD+AI that could be retrieved through the PubMed Central Open Access API. It may not represent prediction model research more broadly; the authors note that this selection could overestimate code sharing compared with the wider literature. The review’s results point to why guidance under development through TRIPOD-Code considers not only whether code is shared, but also how it is documented and organized.