Evaluating Post-Progression Survival in the Context of Progression-Free Survival Benefits: A Revisit of the...

Overall survival (OS) remains the gold-standard endpoint in oncology trials, while progression-free survival (PFS) is a widely used surrogate endpoint that captures tumor progression earlier. However, an increase in PFS has not necessarily led to an improvement in OS (Reck M, 2009). The recent CodeBreaK 200 trial, a randomized, open-label Phase 3 study evaluating sotorasib versus docetaxel as the control arm for previously treated metastatic NSCLC with KRAS G12C mutations (Langen et al., 2023), demonstrated a significant improvement in PFS with sotorasib. However, there was no significant difference in OS between the sotorasib and docetaxel arms, suggesting that PFS may not fully capture the long-term impact of treatment. This discordance led the Oncologic Drugs Advisory Committee (ODAC) to recommend against granting full approval to sotorasib, despite the committee expressing optimism about the drug's potential efficacy (FDA ODAC meeting minutes, 10/05/2023).

In this paper, we aim at identifying the factors that contribute to the discordance between the PFS and OS, especially those that lead to insignificant OS improvement in the presence of the significant benefits to PFS. More importantly, we propose guidance on how to adjust or plan for such factors during the study design in order to better demonstrate the treatment benefit to OS assuming the drug is efficacious.

We proposed to use post-progression survival (PPS), which is defined as the difference between OS and PFS. We developed a flexible parametric simulation framework centered on PPS to quantify how various design factors impact OS given PFS is significant. Based on the simulated data under various scenarios, we investigated the impact of several factors, including censoring informativeness, crossover rates, subsequent therapy access, and improved treatment effect. Moreover, we revisited CodeBreaK200 design as a case study through extracting information from the CodeBreaK200 protocol and digitizing Kaplan Meier curves from published FDA documents.

Our simulations reveal several factors that may lead to the discrepancy between the PFS and OS. First, allowing control-arm patients to crossover to the experimental therapy may lengthen their PPS, thereby shrinking the OS difference between the experimental treatment and control arm. Secondly, permitting both arms to access subsequent treatments similarly narrows the PPS difference. Even a stronger treatment effect that extends PPS in both arms further obscures the OS advantage of the experimental treatment arm. Finally, informative censoring (e.g., due to unbalanced early dropout between the experimental and control arms) may bias estimates of both PFS and OS. When the PFS benefit in the experimental arm is overestimated due to informative censoring, the apparent discordance between PFS and OS may be exaggerated. A larger sample size may help reduce the risk of observing PFS-OS discordance when a true OS benefit exists. In the case study using CodeBreaK200, the combination of informative censoring in the docetaxel arm and post-progression crossover may explain the disagreement between significant PFS benefit and insignificant OS improvement.

We provide recommendations for oncology trial designs that consider factors such as informative censoring, crossover, and subsequent therapies, with the goal of preserving the interpretability of OS outcomes and more accurately reflecting the true treatment effect. An R Shiny application is available to facilitate implementation.