Modeling survival outcomes of KEYNOTE-564 with standard of care control arm treatment: A simulation study.

517 Background: The KEYNOTE-564 study, reporting improved overall survival (OS) with use of adjuvant checkpoint inhibitor (CPI), pembrolizumab, following surgery for renal cell carcinoma (RCC) suffers from a common challenge: a large portion of the control arm received non-standard of care (nSOC) subsequent therapy upon progression (in this case dual CPI or CPI + tyrosine kinase inhibitor [TKI]), thus underestimating survival in this group. We sought to model the survival of KEYNOTE-564 control group patients had they received CPI/TKI per current standard of care (SOC) to more accurately assess this intervention. Methods: Data were extracted from the Kaplan-Meier plots as published in KEYNOTE-426 and KEYNOTE-564. Patient-level survival data were simulated using an iterative process to adjust the censored and event rates to ensure consistency with published reference survival curves derived from KEYNOTE-426. Thus, these measures estimate a counterfactual survival parameter for the KEYNOTE-564 control group based on results from the KEYNOTE-426 study to approximate the expected increase in survival probability if the percent of control patients which received nSOC therapy had received SOC. The analytical process was conducted using R version 4.3.1, with MASS, splines, survival, dplyr, and survminer packages. Results: A total of 52% (N = 109) of patients in the control arm of KEYNOTE-564 that progressed did not receive CPI in the first-line setting. Based on KEYNOTE-426, we estimated a one-year mortality risk of 9.5% for those receiving axitinib-pembrolizumab versus 19.8% in those receiving TKI alone, equating to a 10.3% attributable risk between these arms. When KMA was performed on extracted data where that proportion of patients in the control arm were simulated to have avoided death on an annual basis distributed randomly across five years (n = 51 patients), the re-estimated, unadjusted hazard ratio in this was 1.41 (0.93 to 2.14). As a sensitivity analysis, we conducted this model with a survival parameter of 5% instead of 10% annual mortality risk, resulting in an unadjusted hazard ratio of 0.88 (0.61 to 1.28), and with 2.5% (unadjusted HR 0.72 (0.50 - 1.02). Conclusions: We report and validate method for modeling appropriate subsequent therapy for clinical trials where nSOC treatment in the control group limits the interpretation of their conclusions. In a simulation correcting for the under-estimated survival in the control arm of KEYNOTE-564, the reported survival benefit of adjuvant CPI may not be statistically durable.