Optimizing parametric survival model selection: an analysis using constructed individual patient data extracted from oncology trials from 2014 to 2023

Parametric survival models are widely used in oncology for long-term extrapolation of survival data. We aimed to evaluate fitting and extrapolation performance of parametric survival models and to examine correlation between Akaike and Bayesian Information Criteria (AIC/BIC) and extrapolation accuracy. We included oncology trials supporting Food and Drug Administration (FDA)-approved indications with updated Kaplan-Meier survival curves from 2014 to 2023. Individual patient data (IPD) were reconstructed from survival curves. Standard Parametric Models (SPM), Fractional Polynomials (FP), Restricted Cubic Splines (RCS), Royston-Parmar models (RPM), Generalized Additive Models (GAM), and Parametric Mixture Models (PMM) were evaluated. Goodness-of-fit assessments (extrapolation [GoF-e], a combination of within-sample fit and extrapolation [GoF-fe]) were measured by Mean Absolute Error (MAE) and Restricted Mean Survival Time Difference (RMSD) between parametrically extrapolated and updated Kaplan-Meier curves. Models were compared for their performance in both direct goodness-of-fit and relative goodness-of-fit in preserving treatment effect differences between treatment regimens. Wilcoxon signed-rank test was used for model comparisons. Subgroup analyses by clinical characteristics, and regression analyses were conducted to verify the robustness of results. Spearman tests were used to assess the correlations between within-sample fit and GoF-e. 109 trials providing 346 survival curves were included. AIC and BIC showed no correlations to model extrapolation performance, with Spearman correlations between GoF-f and GoF-e near zero (ρ < 0.1), good within-sample fit did not guarantee reliable extrapolation. RPM consistently demonstrated superior extrapolation performance across both direct and relative goodness-of-fit metrics. For direct GoF-e, compared to SPM (0.446 months/year), median MAE differences were: RPM (-0.032; 95% CI, -0.064~-0.001), PMM (0.097; 0.041 ~ 0.161), FP (0.134; 0.083 ~ 0.191), GAM (0.192; 0.119 ~ 0.271), and RCS (0.246; 0.162 ~ 0.332), RPM significantly outperformed others. For relative GoF-e, RPM also performed the best. Model performance based on GoF-fe was found to be consistent with that based on GoF-e. Sensitivity analyses confirmed the robustness of these findings. Relying solely on AIC/BIC appears insufficient for reliable extrapolation, especially across different classes of parametric models or flexible parametric models. Less-utilized RPM offers the best extrapolation accuracy, has potential for long-term survival predictions.