P46 Effective Use of Reconstructed Survival and Comparative Effectiveness Data: A Case Study from Estimating Unreported Subgroup Survival in Advanced Stage Gastrointestinal Cancers


      This study devises a systematic approach that can utilize aggregate level survival and comparative effectiveness data published from randomized controlled trials (RCT) to assist subgroup-specific health economic and meta-analyses.


      We developed a soft-constrained optimization model, which approximates the restricted mean survival time (RMST) for the overall population in each arm via weighted sum of the RMSTs of two subgroups of interest. Survivals of both subgroups in each arm were assumed to follow Weibull or log-logistic distribution. The constraint ensured that cumulative hazards between the arms were proportional for each subgroup at a sufficiently long pre-specified time point. Estimated subgroup-specific survival functions for the control arm were direct outputs of the model and were shifted by applying the reported hazard ratios from the forest plots to generate their counterparts for the intervention arm assuming proportional hazards between the arms. For validation, we tested our approach in a case study consisting of 10 distinct RCTs with reported subgroup-specific Kaplan-Meier (KM) curves from advanced stage gastrointestinal tumors.


      Across all 48 subgroups, on average, loglogistic model performed equally or better than Weibull model in performance criteria comparing overall survival (OS) rates, median OS and RMSTs. Predicted survival curves laid within the 95% confidence intervals (CIs) of reported KM-curves in 75% and 81% of the time for Weibull and loglogistic models, respectively. Predicted median survivals were within the 95% CIs of the reported medians in 34 and 40 subgroups for Weibull and loglogistic models, respectively. Average relative gap between the predicted and reported RMSTs was 10% in both models. Predicted RMSTs were within the 95% CI of reported RMSTs in 34 and 37 subgroups for Weibull and loglogistic models, respectively.


      Our elicitation approach is effective and demonstrably reliable in deriving unreported subgroup survival with flexible time-varying hazard functions.