reFLASH: a proof-of-concept meta-analytic framework to validate surrogate endpoints in the absence of randomized clinical trial data.

BACKGROUND AND

OBJECTIVES

Within a meta-analytic framework, validating surrogate endpoints traditionally requires individual patient data (IPD) from multiple randomized controlled trials (RCTs). This is rarely feasible in rare and indolent diseases. We propose a proof-of-concept strategy that creates synthetic comparative evidence from multiple single-arm trials to enable trial-level surrogacy assessment.

METHODS

Using IPD from 12 follicular lymphoma RCTs (FLASH; 3996 patients), we decoupled each trial's randomized arms into 24 single-arm trials and re-paired them into 12 artificial trials (reFLASH1) via prespecified rules (therapeutic era, randomization timing, and clinical comparability). Inverse probability of treatment weighting (IPTW) was used to create "comparable" populations between two arms nested within each pseudo trial, to mimic randomization. Marginal treatment effects were estimated using weighted logistic regression for complete response at 30 months (CR30) (odds ratios, OR) and weighted Cox models for progression-free survival (PFS) (hazard ratios, HR). Trial-level surrogacy between CR30 and PFS was quantified using a bivariate mixed-effects (BvME) model and a weighted least squares (WLS) regression model. Robustness was evaluated with leave-one-out (LOO) analyses. A second pairing strategy (reFLASH2) served as a sensitivity analysis. Original FLASH trials (n = 12) were re-analyzed as benchmark.

RESULTS

In reFLASH1, pseudo-trials demonstrated substantial variability in treatment effects (ORs 0.10-15.93; HRs 0.20-2.90), while trial-level surrogacy remained robust, with R2_BvME = 0.769 (95% CI, 0.529-1.000) and R2_WLS = 0.769 (95% CI, 0.341-0.926), meeting predefined acceptability thresholds. In reFLASH2, which applied an alternative arm-matching algorithm, treatment effect ranges were similar (ORs 0.08-8.20; HRs 0.24-6.17), and surrogacy estimates were virtually unchanged (R2_BvME = 0.766; 95% CI, 0.525-1.00; R2_WLS = 0.766; 95% CI, 0.336-0.925). LOO analyses generally kept R2 > 0.70, indicating robustness.

CONCLUSION

Creating artificial comparative trials from single-arm cohorts, combined with IPTW, recovered high trial-level surrogacy between CR30 and PFS, though lower than the RCT-based benchmark obtained under the same analytical framework. While not a substitute for randomization and subject to residual confounding and pairing choices, the reFLASH concept offers a pragmatic pathway to surrogate validation when traditional methods are infeasible due to RCT scarcity.