OBJECTIVES: RCTs often report only relative efficacy measures for subgroups disabling a robust comparison of their long-term survival for subsequent cost-effectiveness analyses. We devised an analytical framework to elicit unreported subgroup specific survival from the aggregate RCT data using lifetime mean survival (LTMS) as an objective.

METHODS: Reconstructed Kaplan-Meier (KM) curves for overall survival (OS) were extrapolated over lifetime using parametric survival models and adjusted with background mortality rates to estimate LTMS in both arms of RCT. By expressing the LTMS in each arm as a weighted average of the mean survival of two select subgroups and assuming exponentially distributed subgroup survival times, hazard rates of the distributions were derived as a closed form solution to two linear equations. Model performance was tested in a case study of 18 distinct RCTs from advanced stage gastrointestinal tumors reporting KM-curves for validation in 96 subgroups in total. For each subgroup predicted median OS (mOS) was compared to 95% CI of the reported mOS. Predictive accuracy of the method was compared to an alternative approach using an objective relying on restricted mean survival time (RMST) limited by the trial follow-up.

RESULTS: LTMS-based approach showed median alignment in 23 subgroups, which was 27 subgroups less than the alignment achieved by RMST-based approach. RMST-based approach had median alignment in 32 subgroups where LTMS-based approach showed none. In 4 RCTs where RMST-based approach showed median alignment in all subgroups, LTMS-based approach showed no alignment in any subgroups. When RMST-based approach allowed subgroup-specific survival times to follow Weibull or loglogistic distribution, median alignment improved by 54 subgroups over LTMS-based approach. Considering splines and dependent models between the arms for survival extrapolations did not improve the performance of LTMS-based approach.

CONCLUSIONS: Eliciting subgroup survival using RMST than LTMS can be more accurate and freer of the uncertainty introduced by long-term survival extrapolations.