EX4 Bayesian MULTI-Parameter Evidence Synthesis to Improve LONG TERM Survival Extrapolation with EARLY Follow up DATA from Nivolumab NSCLC Studies

Abstract

Immuno-oncology therapies can be associated with responses that are delayed but durable and related to long-term survival. This may lead to complex hazard functions which standard parametric models fail to accurately model with often only early trial data available for HTA assessments. We used Bayesian multi-parameter evidence synthesis (B-MPES), a method which enables incorporation of external evidence, to provide long-term survival extrapolations in comparison to standard parametric models. We assessed the performance of B-MPES using databases with minimum follow up of 1- to 5-years from trials evaluating nivolumab versus docetaxel in patients with pretreated metastatic squamous (CheckMate-017) and non-squamous (CheckMate-057) NSCLC. A natural cubic spline model was used to combine overall survival data from each trial with Surveillance, Epidemiology, and End Results (SEER) cancer registry data. Cumulative hazard plots were examined to identify spline knot locations. Performance of the survival extrapolations with 1-, 2-, and 3-year dataset-based B-MPES models were compared with the validated standard parametric model (SPM) for the 5-year dataset using restricted mean survival time (RMST). Additional comparisons against best fit SPMs with earlier data will also be presented. The estimated RMST (95% CI) in months for nivolumab from the 2-, 3-, and 5-year datasets of CheckMate-017 over a 20-year horizon were 24.0 (19.0-29.5), 25.2 (20.7-30.7), and 30.5 (23.0-35.9), respectively, with B-MPES models and 30.5 (21.4-43.6) with the validated 5-year SPM. For CheckMate-057, the estimated 20-year RMST from the 2-, 3-, and 5-year datasets were 22.4 (19.0-26.1), 32.8 (28.2-37.9), and 35.5 months (28.6-44.4), respectively, with B-MPES and 28.6 months (23.5-34.5) with the 5-year validated SPM. We demonstrate that incorporating data from external data sources using B-MPES, yields comparable results to 5-year SPMs using earlier datasets. For CheckMate-057, B-MPES fitted to 3- and 5-year datasets yields more favorable survival extrapolations than the 5-year SPM.