Presentation (CTI)

OBJECTIVES: Survival extrapolation is often a key driver of oncology health technology assessment, but model selection is commonly undertaken before mature overall survival (OS) data are available. This study used sequential OS data cuts from the ALEX trial in NICE technology appraisal (TA) 536 to assess how model plausibility changed with longer follow-up and how extrapolations selected under immature evidence compared with subsequently observed long-term OS.
METHODS: A structured survival model selection algorithm was applied independently to two ALEX trial data cuts: the original immature data cut used in NICE TA536 for alectinib versus crizotinib in anaplastic lymphoma kinase-positive advanced non-small-cell lung cancer, and a later interim data cut with longer follow-up. At each data cut, the algorithm was applied to identify plausible extrapolation models. Model predictions were compared with the final 10-year ALEX analysis. Concordance was assessed descriptively and required predictions for both treatment arms to fall within the observed 95% confidence intervals of OS at 7 and 10 years.
RESULTS: The structured algorithm identified candidate models that outperformed the exponential base-case used in TA536 when validated against subsequently observed 10-year OS. The candidate and plausible model sets changed as trial follow-up matured. At the original data cut, no model family achieved paired-arm concordance with observed OS at both 7 and 10 years. At the later data cut, flexible models became plausible, with the one-knot spline and generalised gamma models concordant in both treatment arms at both time points. The original TA536 exponential base-case extrapolation under-estimated subsequently observed OS in both arms. However, material divergence remained between plausible, better-fitting extrapolations, particularly in the long-term tail.
CONCLUSIONS: Increasing data maturity changed model plausibility, with flexible models becoming plausible and predicting subsequently observed long-term OS more accurately. However, persistent tail uncertainty remained, highlighting the importance of formal methods for incorporating external evidence into survival extrapolation.