MIT and Dana-Farber team uses LLMs to review code sharing in nearly 4,000 medical studies
Only about 1 in 8 clinical prediction studies reported sharing code, and shared repositories varied widely in documentation, the team reports in Nature Medicine.
What happened
Researchers used large language models, the AI behind chatbots, to read nearly 4,000 medical papers. They looked for statements about shared computer code and then checked the code online. Only about 1 in 8 studies said it shared code.
Why it matters to you
Medical prediction models can inform patient care. If other scientists cannot see and run the code, they cannot check how the models were built.
The takeaway
Sharing code is rare in medical prediction research, and shared code is often poorly explained.
What it means for you
Shared code needs enough documentation to understand and run it, and a license file, to be reusable.
Teams building clinical prediction tools can expect journals to set clearer code-sharing expectations as TRIPOD-Code develops.