OBJECTIVES: Cost-effectiveness analyses for health technology assessment require OS predictions over a lifetime horizon. When OS data are unavailable and trial-level relationship between surrogate endpoints and OS cannot be established or is not credible, individual-level surrogacy relationships can be used.

The study objective is to estimate and validate a predictive individual-level surrogacy model between PFS and OS in pre-treated patients with KRAS^G12C-mutated a/m-NSCLC.

METHODS: Patient-level surrogacy of PFS and OS was evaluated based on the joint frailty-copula model described by Emura-2017. Sensitivity analyses considered individual-level frailty factors. The model was fit to PFS and OS data from KRYSTAL-1 Cohort A, a Phase 2 single-arm trial investigating adagrasib in the target population. Covariate selection was based on clinical expert opinion from previous analyses of KRYSTAL-1. Patient-specific OS was simulated using the covariate-adjusted failure function, conditional on progression status at the observed PFS event/censor time. Patients alive when <5% of the sample was at risk of progression were censored, reducing reliance on event-free PFS tails.

The model internal validity was tested by replicating KRYSTAL-1 OS from KRYSTAL-1 PFS. For external validation, the OS of the docetaxel arm of SAPPHIRE, an RCT in patients with pretreated a/m-NSCLC with unknown KRAS mutation, was predicted based on associated PFS. Model performance was measured using restricted mean survival time relative differences (ΔRMST).

RESULTS: In KRYSTAL-1, Kendall’s tau was 0.65 (95%CI: 0.50-0.77), confirming moderate association. The observed OS in KRYSTAL-1 was well-predicted by the model (ΔRMST +1.8%). In the external validity check, model predictions were very similar to the observed SAPPHIRE docetaxel OS (ΔRMST -5.6%).

CONCLUSIONS: The individual-level predictive surrogacy model performed well in predicting OS in pre-treated KRAS^G12C-mutated a/m-NSCLC population, even where the target population differed slightly from the training data. The model predictions will be useful in predicting OS where only PFS data are available.