OBJECTIVES: CAR-T is a type of personalised immunotherapy treatment that has may provide a proportion of people with haematological cancers that are considered difficult to treat, with a potential ‘cure’. A mixture cure model (MCM) is a type of survival model that uses event data from clinical trials to account for cured and non-cured sub-populations within the overall trial population and is considered appropriate for modelling CAR-T outcomes. Immature, highly censored trial data makes it difficult for health technology assessment (HTA) bodies to trust predicted cured and non-cured proportions, however. The objective was to quantify the impact of increasingly mature clinical trial data on the predictive capabilities of MCM.

METHODS: Published ‘ZUMA-1’ trial (axicabtagene ciloleucel for treating refractory large B-Cell lymphoma) Kaplan Meier (KM) data were digitised to generate pseudo individual participant data (IPD). The ‘full’ IPD were truncated to replicate three early trial cuts with increased censoring, representing new datacuts, which were validated with trial publications where available. MCMs were fit to the full dataset and the early trial cuts, using standard parametric distributions for the ‘uncured’ population. Four analyses were compared using graphed long-term survival predictions, goodness-of-fit statistics, median survivals and cure fractions.

RESULTS: Application of MCMs to early trial cuts representing 70% and 60% censoring did not accurately reflect the long-term survival in the full dataset. When censoring is <60%, the outcomes of the early trial cut closely matched the full dataset. There are inconsistencies between goodness-of-fit statistics in all datasets.

CONCLUSIONS: When censoring is >60%, MCMs are not expected to accurately predict long-term survival. When censoring is <60%, the perceived risk that subsequent data collection may change the trial findings is anticipated to be relatively low. Further, with <60% censoring, more consideration should be given to the model capabilities and alternative modelling methods compared to the statistical fit of MCM.