OBJECTIVES:

Lifetime overall survival (OS) extrapolations are required to estimate important differences in costs and benefits between comparators in health economic models. Bayesian multi-parameter evidence synthesis (B-MPES) models for aNSCLC patients receiving second-line nivolumab have previously demonstrated clinically reasonable and stable estimates for long-term OS. Here, we investigate the plausibility of B-MPES predictions for OS in the CheckMate 227 Part 1 trial of nivolumab plus ipilimumab (NIVO+IPI) in aNSCLC patients in the first-line setting and compare with uninformed standard parametric models (SPMs).

METHODS:

The B-MPES approach yields a smooth parametric model that integrates relevant registry data extracted from the Surveillance, Epidemiology, and End Results (SEER) program to inform one-year conditional survival beyond the trial follow-up period. At longer times (>20 years), trial-matched general population mortality data is similarly incorporated to capture an age-related mortality effect. We implemented B-MPES models and SPMs fitted to the NIVO+IPI arm for the 2- and 5-year database locks (DBLs) of CheckMate 227, irrespective of PD-L1.

RESULTS:

The B-MPES model fitted to the 2-year DBL accurately anticipates the observed NIVO+IPI survival in the extended follow-up of the 5-year DBL (5-year OS: 22.4% [95% CrI (credible interval): 19.6-25.5%] vs 22.5% [95% CI (confidence interval): 19.2-26.2%] Kaplan-Meier). The best-fitting 2-year SPM similarly yields agreement with subsequently observed 5-year OS (21.0% [95% CI: 17.8-24.2%]). B-MPES estimates for 20-year OS are conservative and consistent across DBLs, whereas SPM estimates are implausibly optimistic when compared to background mortality data (2-year DBL: B-MPES 1.4% [95% CrI: 1.1-1.7%] vs SPM 5.5% [95% CI: 4.1-7.2%]; 5-year DBL: B-MPES 1.4% [95% CrI: 1.1-1.6%] vs SPM 4.7% [95% CI: 3.6-6.0%]).

CONCLUSIONS:

B-MPES models generate robust estimates of long-term OS for aNSCLC patients receiving NIVO+IPI, even with limited follow-up. SPMs perform reasonably well in anticipating OS observed with extended follow-up but generate unreliable estimates for long-term OS without post hoc adjustment.