OBJECTIVES: In the absence of head-to-head trials, synthesis of clinical effectiveness plays an essential role in health economic evaluations. Interim analyses of trials with limited follow-up are increasingly used to inform regulatory approval so that the evaluations require extrapolating survival curves beyond the immature data. While traditional methods of network meta-analysis (NMA) assume constant treatment effect over whole time horizon, guidance from various health authorities have indicated implausibility of the assumption and suggested incorporation of external data further. This study develops a new extrapolation technique called blended survival curves into the NMA setting to leverage real-world evidence for reliable extrapolation.

METHODS: NMA of blended survival curves involves two separate processes for survival data, in which the first component is a flexible model to fit the network of observed data in the best way possible and the second one is to derive external curves encoding assumptions on the expected behavior of underlying long-term survival. The external process only depends on “hard” evidence, such as registries, cohort studies and clinical opinion. For each intervention, the two survival curves are “blended” into one that is identical to the flexible model over the range of observed times and increasingly similar to the external curve over the extrapolation period. Given blended hazards for all interventions, the time-varying hazard ratios are used to reflect the relative treatment effect.

RESULTS: An example of a network with six trials for first-line treatment of multiple myeloma is used to illustrate the method. Fixed and random effects models were evaluated.

CONCLUSIONS: Long-term extrapolation entirely driven by the less mature data is highly implausible and various assumptions can have a huge impact on the estimate of survival benefits. The blending approach in NMA allows for taking advantage of external data to guide extrapolation and offers significant flexibility without adding substantial computational burden.