OBJECTIVES: Extrapolation of survival outcomes from immature trial data without support from external information may lack justification and be unappealing to payers. To overcome decision uncertainty, we employed Bayesian dynamic borrowing (BDB) to design an informed model for overall survival (OS) for an interim data cut of the KEYNOTE-826 study of pembrolizumab plus chemotherapy with or without bevacizumab (PEMBRO+CHEMO+/-BEV, vs CHEMO+/-BEV) in advanced cervical cancer (aCC), with 15 months minimum follow-up.

METHODS: We formulated a BDB model leveraging historical data from the GOG-240 study of CHEMO+BEV vs CHEMO, with 50 months maximum follow-up. Reconstructed patient-level data from KEYNOTE-826 and GOG-240 were used to estimate generalized gamma survival distributions via BDB, wherein there is a relatively strong a priori preference, imposed via commensurate priors, for KEYNOTE-826 parameter estimates to be similar to concomitantly estimated GOG-240 values. GOG-240 observations were reweighted to reflect the proportion of patients treated with CHEMO+/-BEV in KEYNOTE-826. Projections from Bayesian models with dynamic borrowing and with vague priors were compared to observed OS in the final data cut of KEYNOTE-826.

RESULTS: Leveraging historical trial data for CHEMO+/-BEV led to more conservative and less uncertain estimates of 30-month OS for PEMBRO+CHEMO+/-BEV (40.1% [95% credible interval (CrI): 34.3-46.3%] BDB vs 44.6% [95% CrI: 35.1-51.6%] vague vs 47.4% observed [reconstructed]) and similar but less uncertain estimates for CHEMO+/-BEV (29.6% [95% CrI: 25.6-33.4%] BDB vs 30.2% [95% CrI: 22.3-36.9%] vague vs 31.6% observed [reconstructed]).

CONCLUSIONS: Both BDB and vague Bayesian models forecasted the long-term superiority of PEMBRO+CHEMO+/-BEV vs CHEMO+/-BEV in aCC that was demonstrated in the final data cut of KEYNOTE-826. The BDB method enabled the transparent implementation of conservative assumptions surrounding the longer-term efficacy of the intervention and reduced statistical uncertainty. Payers may consider the BDB approach as a sophisticated method for making informed decisions when presented with immature OS data from oncology studies.