PCN416 NETWORK META-ANALYSIS FOR SURROGATE ENDPOINT EVALUATION IN ADVANCED COLORECTAL CANCER: A BAYESIAN APPROACH

Abstract

Progression free survival (PFS) and tumour response (TR) have been investigated as surrogate endpoints for overall survival (OS) in advanced colorectal cancer (aCRC). Whilst in the past strong surrogate relationship between the treatment effects on TR or PFS and on OS has been found for traditional chemotherapies, in more recent evaluation based on trials investigating modern therapies, the validity of these putative surrogate endpoints has been suboptimal. We aim to evaluate surrogate endpoints by differentiating between therapies of different mechanisms of action. We propose bivariate network meta-analysis (bvNMA) as a method for modelling surrogate relationships between treatment effects on a surrogate and a final clinical outcome within treatment contrasts, borrowing information across treatment contrasts through the network structure of evidence base. We apply this method to data from randomised controlled trials (RCTs) in aCRC. Data were obtained from four systematic reviews of RCTs investigating a range of pharmacological treatments in aCRC, categorised into classes with respect to their mechanism of action. Applying bvNMA to evaluate surrogate endpoints in aCRC resulted in varying correlations between treatment effects on surrogate and final outcome across treatment contrasts. For example overall, for all treatments, correlation between treatment effect on TR and PFS was -0.67 (95% CrI: -0.85, -0.41), whilst the correlation for trials of EGFRi with chemotherapy vs. chemotherapy alone was higher; 0.79 (-0.95, -0.5) and lower for anti-VEGF therapies with chemotherapy vs. chemotherapy alone; -0.43 (-0.84, 0.16). Network meta-analysis allowed us to disentangle information on a relatively strong study-level surrogate relationship between treatment effects on TR and PFS for EGFRi with chemotherapy vs. chemotherapy alone from a set of treatments with suboptimal overall surrogacy relationship. This novel methodology can be used to model surrogate relationships in greater detail compared to methods that do not differentiate between treatment classes.