P-56 Estimating endpoint correlation between surrogate measures and overall survival using reconstructed survival data: Case studies from adjuvant and metastatic gastric cancer trials

Abstract

Validation of intermediate endpoints such as disease-free survival (DFS) and progression-free survival (PFS) as surrogate predictors for overall survival (OS) in randomized controlled trials (RCTs) requires establishing their association at the individual-level. In the absence of individual-level patient data (IPD), this study developed an analytical framework to estimate this association between DFS/PFS and OS using reported Kaplan-Meier (KM) curves from the RCTs and demonstrated its predictive performance in adjuvant and metastatic gastric cancer (GC) treatment settings. Assuming a three-state illness-death model for cancer survival, we developed a linear optimization model to elicit the underlying pre-progression death probability as well as post-progression survival (PPS) distribution using pseudo-patient level DFS/PFS and OS data reconstructed from the published KM curves. In the adjuvant setting, pre-progression death probability was bounded below by the cure rates which were estimated by fitting mixture cure models (MCMs) to the DFS data. In the MCMs, time-to-event outcomes for the uncured subpopulation were modeled using parametric survival functions suggested by National Institute for Health and Care Excellence (NICE) and non-disease-related mortality rates were derived from the age- and sex-adjusted local life-table data from World Health Organization. Reconstructed DFS/PFS distributions were extrapolated via parametric- and spline-based models suggested by NICE and adjusted with estimated background mortality rates whereas elicited PPS distributions were extrapolated assuming constant hazard rate over time. Estimated pre-progression death probabilities and modeled DFS/PFS/PPS distributions governed a Monte-Carlo simulation framework which generated paired pseudo pre- and post-progression data to predict Spearman’s rank and Pearson’s product moment correlation coefficients. Model performance was tested on two correlation meta-analyses in GC (14 RCTs with 3371 patients on adjuvant chemotherapy; 20 RCTs with 4069 patients on metastatic treatments) published in 2013 by the GASTRIC group. For each test case, model-predicted OS rates and Spearman rank correlation coefficients were compared against their reported counterparts and corresponding 95% CIs. Predicted OS curves laid within the 95% CIs of the reported OS KM-curves 96% and 100% of the time in the adjuvant and metastatic setting, respectively, where the average deviation between the restricted mean survival times under the model-predicted OS curves and the statistically best-fitting OS curves to the reported data was < 1% in both settings. Average deviation between the estimated and reported Spearman rank correlation coefficients was no more than 0.01 (reported: 0.97 [95% CI:0.97-0.98] vs. predicted: 0.96 [95% CI:0.96-0.96]) and 0.13 (reported: 0.85 [95% CI:0.85-0.85] vs. predicted: 0.72 [95% CI:0.71-0.72]) in both settings. Predicted Pearson correlation coefficients were 0.95 [95% CI:0.95-0.95] and 0.94 [95% CI:0.94-0.95] in the adjuvant and metastatic setting, respectively. Our study offers a useful approach for an indirect endpoint correlation assessment in the absence of IPD. Results indicate the model to be precise in adjuvant but conservative in metastatic GC setting which should be approached with caution due to independent simulation of paired DFS/PFS and PPS durations from the illness-death model and the lack of data-driven lower bounds on pre-progression death probability in the metastatic setting.