An alternative method to validate surrogate endpoints in oncology

Surrogate endpoints are often utilized to support drug efficacy claims. Hence, it is of paramount importance to validate the effectiveness of surrogate endpoints in correctly predicting clinical benefit, in order to ensure a fast and safe access to new drugs. This study aimed to explore the applicability of Receiver Operating Characteristic (ROC) curve and its Area Under Curve (AUC) for trial-level surrogate validation, and to compare the predictive power of surrogate threshold effects (STE) derived from AUC versus linear regression for survival benefits. Based on a meta-analysis of randomized controlled trials of advanced gastroesophageal cancer, we extracted the treatment effect data on surrogate endpoints (objective response rate (ORR) or progression-free survival (PFS)) and overall survival (OS). Treatment effects on OS were dichotomized as statistically significant improvement and others. The ROC curve and its AUC were obtained by calculating the sensitivity and specificity of OS status at each surrogate treatment effect cutoff value. As a contrast, the linear regression was modeled to derive the determination coefficients ($$\:{R}^{2}$$) between OS and surrogates. STEs were estimated using AUC and linear regression, respectively, and their performance in correctly predicting OS benefit status was analyzed and compared. A total of 87 trials were collected. Both approaches showed that PFS had better quality than ORR ($$\:{R}^{2}$$: 0.45(95% CI, 0.30–0.61) vs. 0.21(95% CI, 0.1–0.36); AUC: 0.84(95% CI, 0.75–0.93) vs. 0.71(95% CI, 0.59–0.83)). The AUC-based STE performed substantially better in predicting significant OS benefit than the linear regression-based STE for both ORR (sensitivity: 81.3% vs. 0%; accuracy: 62.2% vs. 54.1%) and PFS (sensitivity: 88.9% vs. 41.2%; accuracy: 79.5% vs. 71.8%). The ROC curve and its AUC are built on a binary indicator of the statistical significance of final-outcome treatment effect, thereby improving the accuracy of predicting true efficacy superiority. This approach demonstrates a potential in evaluating the ability of a surrogate endpoint to predict final-outcome benefits.