OBJECTIVES: Subgroup analyses are crucial for health technology assessments, yet indirect comparisons and economic evaluations for subgroups often rely on limiting assumptions. We developed a novel analytical framework to elicit unreported subgroup-specific Kaplan-Meier (KM) curves and underlying cumulative event data from aggregate randomized controlled trial (RCT) data.

METHODS: Reconstructed KM-curves from the two arms of an RCT, and reported hazard ratios (HRs) between two exclusive and exhaustive subgroups were used to formulate a linear optimization model. Decision variables represented patients’ likelihoods of belonging to either of the subgroups in each arm. Model was parametric in unreported cumulative number of events for each subgroup in each arm, and repeatedly solved for all possible combinations with the objective of minimizing the gap between reported and predicted HRs for both subgroups. Ties among multiple optimal solutions were broken by comparing predicted and reported 95% CIs for subgroup-specific HRs. Model performance was tested in a comprehensive case study including 30 distinct RCTs (18 metastatic, 12 adjuvant) from gastrointestinal tumors reporting KM-curves for overall survival or an intermediate outcome in 176 (96 metastatic, 80 adjuvant) subgroups. For each subgroup, predicted number of events, survival rates and restricted mean survival times (RMSTs) were compared against actual values and their 95% CIs.

RESULTS: Across all subgroups, average absolute gap between predicted and reported monthly survival rates was less than 5%. Predicted survival curves laid within the 95% CIs of the reported survival curves in 83% (metastatic) and 79% (adjuvant) of the time. Predicted RMSTs were within the 95% CIs of their reported counterparts in 68 (metastatic) and 63 (adjuvant) subgroups. Average relative gap between predicted and reported number of deaths across all subgroups was 8.5% (metastatic) and 2.3% (adjuvant).

CONCLUSIONS: The proposed approach generalizes previously published distribution-free framework eliciting unreported subgroup-specific survival in a scalable fashion without loss of accuracy.