MainClinical prediction models now inform diagnostic and prognostic decisions across a wide range of care settings. Before such models can be trusted in practice, their validity must be independently assessable, their performance must be reproducible in different cohorts to establish generalizability and their development must be transparent enough to support translation into real-world deployment. The analytical code that is used to, for example, preprocess data, develop the prediction modeling approach, select or tune model parameters and evaluate performance is essential for methodological assessment and reproduction of results. Yet access to the code underlying published prediction models remains uncommon1.Concerns about computational transparency in scientific research are longstanding, and recent initiatives have sought to strengthen reporting and sharing practices. The FAIR principles have outlined standards for reusable digital research artifacts2; the EQUATOR Network has advanced reporting guidelines to enhance methodological transparency3; and journals and funders increasingly require data- and code-availability statements to promote reproducibility4,5,6.In clinical prediction research, the TRIPOD Statement and its 2024 update, TRIPOD+AI (www.tripod-statement.org)7,8,9, were developed to improve the completeness and clarity of model reporting. These guidelines have helped promote methodological quality, external validation and calibration of prediction model research10. However, although TRIPOD+AI encourages sharing of analytical code, it does not explicitly specify standards for repository structure, documentation or reproducibility. Consequently, code-availability reporting is likely to be heterogeneous11, although the actual extent and quality of code sharing in prediction model research have not yet been studied.Evidence from the broader biomedical literature indicates that code sharing is infrequent and that declared code availability often does not correspond to functional accessibility12,13. In a large-scale attempt to rerun Jupyter notebooks linked to biomedical articles in PubMed Central, prior work identified thousands of Python Jupyter notebooks but found that only 5.3% could be run successfully end-to-end14. Manual review of code repositories reports similar gaps: among articles accepted at the Medical Imaging with Deep Learning (MIDL) conference, only 22% had a public repository judged ‘repeatable’ under predefined reproducibility criteria15, and in a study evaluating 160 deep-learning computational pathology articles, only about one-quarter had made code publicly available16.However, these investigations focused on specific code formats (for example, Jupyter notebooks) or assessed code documentation through a labor-intensive manual review that does not scale to large, diverse corpora. Prior work has shown the potential of large language models (LLMs) to facilitate large-scale reviews17,18, but to our knowledge, no study has assessed code sharing and repository documentation practices in prediction model research, applied LLMs to jointly extract code-sharing statements from articles or characterized repository features at scale. Establishing this empirical baseline is necessary to understand the current practice of code sharing in prediction model research and to inform the development of more precise reporting standards.In response to this need, we are developing the TRIPOD-Code extension to provide structured guidance on code availability and computational reproducibility in prediction model research19. Conforming to the original TRIPOD definition, we considered a multivariable prediction model as any combination of multiple predictors used to estimate an individual’s probability of the presence of a particular health condition (diagnostic) or whether a particular outcome will occur in the future (prognostic). The present study represents Stage 1 of the project: a large-scale, scoping review of studies citing TRIPOD or TRIPOD+AI, undertaken to quantify code availability and evaluate the characteristics of shared repositories.ResultsWe identified 6,762 PubMed articles citing the TRIPOD Statement and 411 citing the TRIPOD+AI Statement as of 11 August 2025 (Fig. 1). After removing 515 duplicate entries, the initial cohort comprised 6,658 unique articles with associated PubMed metadata. Of these, 1,407 articles were not retrievable through the PMC Open Access API, and an LLM screening process found that 1,284 articles did not develop, update or validate a model that combined information from multiple predictors or input features to estimate the probability of the presence or future occurrence of a particular outcome. This left 3,967 papers that met our selection criteria (see Supplementary Data 1).Fig. 1: Modified PRISMA flow diagram.Full size imageWe determined that an article was sharing code (n = 482) if: it shared code in the appendix (n = 75), its repository provider was not supported (n = 27) or its repository was included in the review (n = 380).Code sharingAmong the 3,967 articles included in the review, 12.2% (n = 482) shared their code. Code was most frequently shared through a repository link, whereas 75 articles reported code availability in an appendix or supplementary material. Code sharing increased over time across the study period (Fig. 2), with the odds of code sharing increasing by 12.4% per publication year in binary logistic regression (odds ratio (OR), 1.12; 95% confidence interval (CI), 1.07–1.18; P