Presentation (CTI)

OBJECTIVES: Overall survival (OS) is the gold standard endpoint but often requires long follow-up, while disease-free survival (DFS) is frequently used as a surrogate endpoint. However, meta-analytic surrogacy evaluations typically rely on trial-level associations because individual patient data (IPD) are rarely available, precluding direct modeling of individual-level dependence. We propose a two-stage Bayesian meta-analytic framework that reconstructs pseudo-IPD from published survival curves to infer individual-level dependence and incorporate it into surrogacy evaluation and prediction.
METHODS: We developed a two-stage framework linking surrogate and target endpoints by combining individual- and trial-level evidence. First, DFS and OS Kaplan-Meier curves were digitized from HER2 positive early-stage breast cancer trials and used to reconstruct pseudo-IPD. A rank-based pairing algorithm with a tunable σ parameter was applied to calibrate individual DFS-OS pairs. Within each trial, the DFS-OS correlation (ρwi) was characterized using three approaches: (1) a non-informative uniform prior on (-1,1); (2) residual-based correlations from Cox proportional hazards models fitted to pseudo-IPD; and (3) individual-level dependence estimated via a Plackett copula and transformed to ρwi. Second, trial-level treatment effects for DFS and OS were summarized as log hazard ratios (HRs) with corresponding standard errors and analyzed within the Bayesian Daniels and Hughes meta-analytic framework, incorporating trial-specific ρwi estimates from the first stage. Performance was evaluated based on predictive accuracy and model fit.
RESULTS: Incorporating individual-level correlation information reduced uncertainty around predictions of OS HRs. Among the approaches evaluated, the copula-based method yielded the most precise estimates and superior predictive performance for OS treatment effects based on DFS compared with alternative correlation specifications.
CONCLUSIONS: Incorporating within-study correlation, particularly through copula-based methods, improves the precision and reliability of surrogacy evaluation and OS prediction. This two-stage Bayesian framework is tractable without individual patient data, relying on published survival curves and trial-level summaries, and supports robust surrogacy assessment for health economic and regulatory decision-making.