Fairness auditing of machine-learning and LLM stroke-risk estimation: analysis code and intermediate predictions (NHANES 2007-2018 development, CHARLS 2011 external validation)
Analysis code, development-cohort intermediate predictions, and machine-readable results supporting the manuscript "Fairness auditing of machine-learning and LLM stroke-risk estimation: development in NHANES 2007-2018 and external validation in CHARLS 2011". Contents: (1) full analysis pipeline scripts covering cohort construction, four-model training with subgroup fairness audits (LR/RF/GBM/MLP plus reweighing and adversarial debiasing), zero-shot LLM vignette audits of Qwen2.5-1.5B/7B, external validation in CHARLS 2011, and QLoRA distillation; (2) out-of-fold predictions for all tabular models on the NHANES development cohort (n = 34,719),
Resource information
- Format
- zip
- Original repository
- 22725277
- DOI
- 10.5281/zenodo.22725277
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- cc-by-4.0
- Last record update
- Nov 5, 2026
About this resource
Analysis code, development-cohort intermediate predictions, and machine-readable results supporting the manuscript "Fairness auditing of machine-learning and LLM stroke-risk estimation: development in NHANES 2007-2018 and external validation in CHARLS 2011". Contents: (1) full analysis pipeline scripts covering cohort construction, four-model training with subgroup fairness audits (LR/RF/GBM/MLP plus reweighing and adversarial debiasing), zero-shot LLM vignette audits of Qwen2.5-1.5B/7B, external validation in CHARLS 2011, and QLoRA distillation; (2) out-of-fold predictions for all tabular models on the NHANES development cohort (n = 34,719), per-vignette LLM audit outputs (n = 1,000), and QLoRA per-case predictions; (3) QLoRA adapter weights for both models; (4) JSON results for each analysis part. NHANES-derived files are redistributable (public domain). CHARLS microdata are not redistributed and are available from the CHARLS team (https://charls.charlsdata.com/) under a data use agreement.