OBJECTIVES: Oncology trials are typically powered on progression free survival (PFS). At time of health technology assessment the overall survival (OS) is therefore often immature. Mixture cure models need a sufficiently long survival tail to capture a cured fraction. NICE TSD guidance on extrapolating survival data indicates that external data may be used to reduce uncertainty. The aim was to inform overall survival cure rates of the active and placebo arm by jointly modelling PFS and OS with a mixture cure model.

METHODS: Bayesian loglogistic cure models were defined where individual cure rates of active and placebo were jointly estimated for PFS and OS. The shape and scale parameters of the loglogistic distribution were modelled separately for PFS and OS by treatment. In a separate analyses, the posterior PFS cure rates estimated with the loglogistic MCM were used as prior for the loglogistic MCM on OS data. These two approaches were compared with the cure rates derived when considering the overall survival in isolation with a MCM loglogistic. A trial in front line renal cell cancer comparing ipilimumab plus nivolumab with sunitinib was used as example.

RESULTS: The loglogistic MCM over PFS showed cure rates of 0.14 [0.08;0.20] and 0.33 [0.26;0.39] for sunitinib and ipilimumab plus nivolumab respectively. For OS the corresponding numbers were 0.21 [0.11;0.30] and 0.26 [0.06;0.41]. The Bayesian loglogistic MCM jointly estimating cure rates for PFS and OS showed cure rates of 0.14 [0.09;0.20] and 0.33 [0.28;0.38]. The OS MCM using PFS cure rates as prior, showed cure rates of 0.14 [0.09;0.20] and 0.32 [0.21;0.42].

CONCLUSIONS: The presented methods can inform overall survival predictions using mixture cure models causing reduced uncertainty in predictions, while still being in the confidence range of the uninformed MCM for OS. This is in line with NICE TSD guidance. Assumptions need to be validated with clinicians.