Stratified Covariate-Reconstructed Survival (SCRS): Extending Guyot IPD Reconstruction to Preserve Covariate-Outcome Associations for Population- Adjusted Indirect Comparisons

Abstract Background Reconstructing individual patient data (IPD) from published Kaplan–Meier curves using the Guyot algorithm is widely used in health technology assessment and network meta-analysis. However, the standard algorithm produces time-to-event data without patient-level covariates, precluding population-adjusted indirect comparisons such as matching-adjusted indirect comparison (MAIC). We propose Stratified Covariate-Reconstructed Survival (SCRS), an extension that preserves covariate-outcome associations in reconstructed IPD through stratified assignment and within-stratum rank-based coupling. Methods SCRS operates in four phases: (1) Guyot reconstruction of Kaplan–Meier curves; (2) virtual cohort generation matching published baseline characteristics via Gaussian copulas; (3) stratified assignment of survival times with adaptive rank-based coupling, ensuring high-risk patients within each stratum receive earlier event times; and (4) multiple imputation across B reconstructions with Rubin’s rules for pooled inference. We validated SCRS using the ASPIRE pulmonary arterial hypertension registry (N = 773, true IPD) and conducted Monte Carlo validation across 1,300 synthetic trials spanning 13 scenarios (including oncology and cardiology) that varied stratification richness (1–3 variables), MAIC dimensionality (2–6 balancing variables), event rates (8–60%), and sample sizes (80–500). Three reconstruction methods were compared: naive random assignment, stratified random assignment, and rank-coupled stratified assignment. Results In ASPIRE validation (B = 50 imputations), rank-coupled SCRS recovered 83% of discriminative capacity (C-index recovery) versus 55% for stratified random and 10% for naive assignment, and recovered 80% of stratification-variable hazard ratios versus 56% for stratified random. In Monte Carlo validation across 1,300 synthetic trials spanning three disease areas (PAH, oncology, cardiology), rank-coupled SCRS recovered a median 101% of discriminative capacity versus 71% for stratified random and 27% for naive, outperforming stratified random in 89% of trials (95% CI: 87–91%). Cross-disease generalizability was confirmed: win rates ranged from 87% (PAH) to 99–100% (oncology, cardiology). MAIC root mean squared error was reduced modestly (171 vs 173 for stratified random, 218 for naive). Calibration RMSE at 60 months improved by 7% (0.057 vs 0.061) in ASPIRE validation. Conclusions SCRS extends standard IPD reconstruction to enable population-adjusted indirect comparisons from published data. Rank-based coupling within strata nearly fully recovers discriminative capacity (96% median in Monte Carlo, 83% in ASPIRE), a substantial improvement over stratified random assignment alone (71% and 55% respectively). The method is recommended for MAIC-based population adjustment when cross-trial heterogeneity is substantial, published subgroup data support stratification, and the arm of interest contains at least 10 events. Due to within-stratum hazard ratio attenuation (10–20% for stratification variables, 30–50% for unstratified variables), the method is better suited for MAIC (weighting-based) than STC (outcome regression-based) population adjustment.