Comparative evaluation of graph attention for survival prediction and prognostic stratification in non-small-cell lung cancer: supporting reproducibility materials

This record provides supporting reproducibility materials for the study Comparative evaluation of graph attention for survival prediction and prognostic stratification in non-small-cell lung cancer. The study retrospectively analyzed 112,345 patients with stage I–IV non-small-cell lung cancer from the Surveillance, Epidemiology, and End Results (SEER) Program. The development cohort comprised 46,762 patients diagnosed during 2018–2019, and the temporal validation cohort comprised 65,583 patients diagnosed during 2020–2022. The archive includes aggregate study tables, model configurations, a 48-feature clinical indicator dictionary, supporting performance and sensitivity analyses, and reference Python code covering feature preparation, survival modeling, graph construction, normalization, prognostic-group assignment, and evaluation. The study compared stage-only Cox regression, multivariable Cox regression, XGBoost Cox, multilayer perceptron Cox, and graph attention network Cox. Checks using privately retained prediction outputs reproduced the reported temporal concordance values for XGBoost Cox (0.7665), multilayer perceptron Cox (0.7630), and graph attention network Cox (0.7516). The reference code reconstructs the described analytical methods but has not been verified as the original final training implementation. Independent reproduction of the complete training pipeline has not been established. The precise original SEER extraction configuration and complete software environment remain unverified. No individual-level SEER records, patient-level predictions, neural embeddings, or trained model parameters are distributed. Researchers requiring individual-level SEER data must obtain access through the applicable SEER procedures. A development-cohort aggregate table has been abridged as a conservative disclosure measure. Figures are not included. The materials are intended to support methodological transparency, interpretation of the reported results, and further research into clinical survival prediction and graph-derived prognostic stratification.