Flexible Survival Extrapolation with Blended Hazards: Accounting for Treatment Effect Waning in Health Technology Assessment
Jingqi Zhu, Matthew Hemstock, Zhaojing Che, Gianluca Baio, Richard BirnieBackground
Survival extrapolation is crucial in estimating the lifetime survival benefit of a treatment for health technology assessment (HTA). Conventional extrapolation methods, which assume that the long-term treatment effect (hazard ratio between treatment and comparator) follows the same pattern as observed in the short-term trial, have been challenged by a wide range of immuno-oncology therapies, particularly those with administrative stopping rules that mandate treatment discontinuation at a prespecified time point. A gradual waning of their treatment effects has been considered plausible and received growing attention from HTA stakeholders over the past decade. However, existing statistical methods often rely on unnecessarily strong waning assumptions.
Objective
We demonstrate the blended hazard method as a flexible way to account for treatment effect waning while incorporating external evidence in survival extrapolation.
Method
The blended hazard method fits separate parametric survival regression models to the observed randomized controlled trial data and external data that inform the common long-term hazard when there is no treatment effect. For each arm, the fitted internal and external hazard functions are blended based on a time-varying weight function, so that the blended hazard is initially dominated by the fitted internal hazard, then gradually approaches the fitted external hazard over a blending interval, and is finally dominated by the fitted external hazard. The time and rate of blending the internal and external information can be controlled by the weight function to allow for sensitivity analyses. NICE TA366 on pembrolizumab for advanced melanoma not previously treated with ipilimumab is used as a case study to demonstrate the practical implementation of this method.
Results
Extrapolations and restricted mean survival times from the blended hazard method closely matched the updated 7-y trial follow-up and showed better consistency than the TA366 base case across all sensitivity analysis scenarios.
Conclusion
The method explicitly accounts for gradual treatment effect waning while incorporating external evidence and offers the flexibility to accommodate a broad range of waning scenarios, thereby effectively characterising uncertainty in extrapolation.
Highlights
Treatment effect waning is considered plausible in survival extrapolation for many therapies, particularly those with treatment-stopping rules. However, there is a shortage of appropriate methods to model this phenomenon, and existing approaches either rely on strong waning assumptions or address it only as a post hoc check.
We demonstrate the blended hazard method as a possible approach to account for treatment effect waning while incorporating external evidence.
The blended hazard method possesses the flexibility to accommodate a wide range of waning scenarios, thereby relaxing unnecessarily strong assumptions and effectively characterizing uncertainty in survival extrapolation.