OBJECTIVES: Immunotherapies have the potential to yield durable response and may therefore lead to complex lifetime hazards, which motivates leveraging external data to generate more reliable survival extrapolations. Bayesian methods provide a holistic framework to incorporate external data into parametric survival models, but some formulations require additional inputs that may be difficult to choose a priori. Here, we investigate Bayesian poly-hazard models, which have attractive features of flexibility, straightforward clinical interpretation, and absence of mandatory user-specified auxiliary parameters. We apply the approach to overall survival (OS) data for nivolumab plus ipilimumab (NIVO+IPI) in first-line aNSCLC from CheckMate 227 Part 1, with 29.3 months of minimum follow-up.

METHODS: We consider a Bayesian poly-hazard model with background and disease-specific components, for which the corresponding prior distributions are estimated from general population and relevant SEER registry data, respectively, using a fixed background hazard in the latter step. Model predictions are compared to 5-year observations from extended follow-up and to previously reported estimates from Bayesian multi-parameter evidence synthesis (B-MPES) and an uninformed standard parametric model (SPM).

RESULTS: Short-term extrapolations from the Bayesian poly-hazard model agree with later observations (5-year predicted OS: 20.7% [95% CrI (credible interval): 17.7-23.8%] vs 22.5% [95% CI (confidence interval): 19.2-26.2%] Kaplan-Meier), even though the SEER data represents a highly pessimistic expectation for NIVO+IPI survival. Long-term extrapolations are conservative compared to the naïve SPM and are in close agreement with B-MPES under a pessimistic scenario (20-year OS: 1.8% [95% CrI: 1.2-2.6%] poly-hazard vs 5.5% [95% CI: 4.1-7.2%] SPM vs 1.4% [95% CrI: 1.1-1.7%] B-MPES). The poly-hazard model predicts that the background and disease-specific hazards become equal at approximately 15 years.

CONCLUSIONS: Bayesian poly-hazard models provide a flexible approach with clinically interpretable structure to plausibly estimate lifetime conditional survival for patients with aNSCLC receiving NIVO+IPI. Moreover, their data-driven formulation helps to reduce subjectivity in survival extrapolations.