Predicting Overall Survival Benefit in Unresectable or Metastatic Melanoma from Improvement in Progression-Free Survival: A Correlation Meta-Analysis

Peter Mohr^1*Murat Kurt²Swetha Srinivasan²Andriy Moshyk²Flavia Ejzykowicz²Paul Serafini³Mir-Masoud Pourrahmat³Lisa Leung³

• ¹Department of Dermatology, Elbe Kliniken Buxtehude, Buxtehude, Germany
• ²Bristol Myers Squibb, Princeton, NJ, United States
• ³Evidinno Outcomes Research Inc., Vancouver, Canada

ObjectivesTo evaluate the association between the treatment effects on progression-free survival (PFS) and overall survival (OS) for previously untreated, unresectable or metastatic melanoma.MethodsA systematic literature review identified eligible trials reporting PFS and OS. Bivariate random effects meta-analysis (BRMA) was performed to estimate the correlation between the hazard ratios (HRs) of OS (HROS) and PFS (HRPFS), and sample size-weighted linear regression (WLR) was used to estimate a surrogacy equation which predict the HROS from the HRPFS. Strength of the correlation obtained from BRMA and WLR models was assessed using published guidelines. Predictive performance of the WLR model was also evaluated internally by leave-one-out cross-validation (LOOCV) and externally against data from newly published trials. Further analyses included adjustments for BRAF mutation status, and restriction to phase III trials or trials evaluating immune checkpoint or BRAF/MEK inhibitors, without crossover or crossover-adjusted, or meeting proportional hazards assumption.ResultsBRMA and WLR estimated a correlation of 0.74 (95%CI: 0.51-0.87) and 0.81 (95%CI: 0.58-0.92), respectively. The estimated surrogacy equation derived from the WLR was lnHROS = -0.05 + 0.50 × lnHRPFS with a statistically non-significant intercept (95% CI: -0.14 - 0.03) and a statistically significant slope (95% CI: 0.35 - 0.65). The surrogacy equation derived from the BRMA was lnHROS = -0.11 + 0.36 × lnHRPFS with a statistically non-significant intercept (95% CI: -0.23 - 0.00) and a statistically significant slope (95% CI: 0.17 - 0.57). The predictive accuracy of the WLR was 95.8% in LOOCV. Across sensitivity analyses correlations between HRPFS and HROS were ≥0.77 and ≥0.85 based on BRMA and WLR, respectively, and the accuracy of the WLR model in LOOCV was ≥88%. When predicting HROS for newly published trials, the differences between the observed and model-predicted HROS’s were <0.05.ConclusionsResults suggest a clinically meaningful and moderate trial-level correlation between PFS and OS across all analyses. The analyses and high accuracy of the surrogacy equations shown in internal and external validations can enable earlier prediction of treatment effects on OS from the improvements on PFS for previously untreated unresectable or metastatic melanoma.