OBJECTIVES: Survival extrapolations from naïve standard parametric models (SPMs) can be highly unreliable when fitting to early cuts of study data, even for short-term projections. Utilizing external data is therefore desirable, but predictions can then be dependent on the data sources and implementation method. Here, we extend the Bayesian multi-parameter evidence synthesis (B-MPES) framework to incorporate historical trial data and fit a B-MPES model to 1-year data from CheckMate 9LA, which compares nivolumab plus ipilimumab plus two cycles of chemotherapy (NIVO+IPI+CHEMO) versus four cycles of CHEMO in first-line mNSCLC.

METHODS: To illustrate the approach, we estimate a B-MPES model utilizing observations for NIVO in second-line advanced NSCLC from the 5-year data cuts of CheckMate 017 and 057 to conservatively inform NIVO+IPI+CHEMO survival predictions beyond the CheckMate 9LA follow-up period. This historical trial data is not necessarily ideal but is intended to provide information on the potential for longer-term survivorship with immunotherapy, since this phenomenon is not represented in the immature trial data. Longer-term extrapolations are further supported by SEER registry and general population data. B-MPES estimates are compared to extrapolations from an uninformed SPM and 4-year observations from extended follow-up of CheckMate 9LA.

RESULTS: The B-MPES model fitted to the 1-year data yields predictions that are highly consistent with subsequently observed NIVO+IPI+CHEMO survival (4-year survival: 19.9% [95% CrI (credible interval): 16.7-23.2%] B-MPES vs 21.0% [95% CI (confidence interval): 17.2-25.6%] Kaplan-Meier). In contrast, short-term extrapolations from the 1-year uninformed SPM are more conservative (4-year survival: 17.9% [95% CI: 14.5-21.3%]). At longer times, B-MPES is more pessimistic than the SPM (20-year survival: 1.2% [95% CrI: 0.9-1.5%] B-MPES vs 2.1% [95% CI: 1.5-3.0%] SPM).

CONCLUSIONS: When relevant historical trial data are integrated into a survival model appropriately, this information can improve the reliability of short- and long-term extrapolations and is especially useful when trial data are immature.