OBJECTIVES: Novel oncology treatments often undergo health technology assessments based on immature trial data due to high unmet need. Extrapolation of overall survival (OS) with parametric survival models (PSM) can be subject to large uncertainty in these situations. More mature OS data from prior studies in the same indication and with similar mechanism of action can be used to derive informative priors (IPs) for long-term OS extrapolation. This case study in newly diagnosed advanced melanoma (aM) investigates whether Bayesian PSM (bPSM) using IPs provide better extrapolations compared with standard PSM.

METHODS: The CoBRIM trial was used for the case study with long term OS serving as the reference for analyses (i.e., true distribution to be estimated). Reconstructed individual patient level data for the CoBRIM trial were generated and truncated to create immature data cuts (median follow-up 9.5 months vs 21.2 months with complete observation). Published data from the Combi-V, Combi-D, BRF120330 and DREAMseq trials were used to generate IPs for the ancillary parameters of commonly fitted PSM distributions. Standard PSM and bPSM using these IPs were then fitted to CoBRIM immature data and predicted restricted mean survival times (RMST) from the best fitting distribution were compared to those derived from the reference OS distribution.

RESULTS: The lognormal distribution achieved the best fits for PSM and bPSM. Estimated RMST for OS using bPSM (32.2 months) was much closer to the ground truth (35.3 months) than the PSM estimate (28.9 months). At 24, 36 and 48 months ground truth OS was 48.6%, 38.6% and 34.4% , respectively, which compared better with bPSM extrapolations (49.9%, 36.4%, 27.7%) as opposed to standard PSM (45.0%, 30.3%, 21.3%). Results were consistent with alternate IPs.

CONCLUSIONS: Use of bPSM can improve extrapolation of immature OS data when suitable IPs can be formulated from prior studies.