PCN292 BAYESIAN NETWORK META-ANALYSES OF OVERALL SURVIVAL: APPLICATION OF DIFFERENT PARAMETRIC AND FRACTIONAL POLYNOMIAL METHODS

Network meta-analysis (NMA) of time-to-event outcomes often relies on pooling of hazard ratios (HR). However, in immuno-oncology, there is increasing evidence of time-dependent treatment effects and violation of proportional hazard (PH) assumption within trials. We performed a multidimensional NMA of trials comparing PD-1/PD-L1 inhibitor plus chemotherapy with chemotherapy in the first line of treatment for advanced squamous non-small cell lung cancer. Two studies were identified in an existing systematic review (Zhou, 2018) evaluating pembrolizumab and atezolizumab in combination with chemotherapy vs chemotherapy. The Kaplan Meier curves for intervention and comparator arms were “digitized” using DigitizeIt. The PH assumption between the survival curves was violated, and a comparison of constant HRs seems inappropriate. The Bayesian NMA was performed using a series of parametric and fractional polynomial (FP) models. Fixed effect models were fitted to explore long-term survival. The parameters of the different models were estimated using suvivalNMA package in R (v3.6.2). Deviance information criteria (DIC) was used to compare the goodness-of-fit. Weibull and FP of order 1 showed the best fit with comparable DIC (1020.4 and 1020.9, respectively). The predicted 2 years and 5 years survival of pembrolizumab, atezolizumab and chemotherapy were similar with both best-fitted distributions, i.e. 47%, 42% & 31% for 2 years and 10%, 8% and 2% for 5 years respectively. Pembrolizumab plus chemotherapy showed higher overall survival, followed by atezolizumab plus chemotherapy and chemotherapy. The findings were not sensitive to different distributions, although the survival probabilities varied substantially. The results were aligned with the coxPH indirect treatment comparison. Network meta-analysis of time-to-event outcomes using multidimensional parameters can be more closely fitted to data over analysis based on constant HRs. Use of this approach may be more informative for any subsequent cost-effectiveness modelling over constant HRs.