Presentation (CTI)

OBJECTIVES: Survival extrapolation in oncology is essential to inform cost-effectiveness in health technology assessment (HTA). HTA guidelines recommend incorporating external evidence to improve long-term plausibility, but how this should be operationalised, and how robust the estimates are, remains poorly characterised. This study assessed whether integrating external trial data into a Bayesian framework improves long-term overall survival (OS) accuracy versus standard parametric survival models (PSMs).
METHODS: Clinical trials assessing melphalan-prednisone (MP) in transplant-ineligible newly-diagnosed multiple myeloma were identified through a literature review. OS from VISTA (75-month follow-up; restricted mean OS 42.4 months) was truncated at 30 months to emulate immature data. Seven standard PSMs were fitted to the truncated data. Bayesian PSMs (BPSMs) were then fitted using informative priors on shape parameters from eight external MP trials, under three prior strengths (weak: SD=20% of mean; moderate: 5%; strong: 0.1%). Accuracy was quantified as the area between each extrapolation and the full Kaplan-Meier curve in months (ΔAUC).
RESULTS: Standard PSMs overestimated long-term OS, with ΔAUC of 1.9 to 4.5 months (4.5% to 10.6% of restricted mean survival). In contrast, BPSM extrapolations were distributed around the Kaplan-Meier curve rather than systematically above it. Under moderate informative priors, adding an external prior improved accuracy versus the same distribution without a prior in 83% of cases. Moderate priors outperformed weak priors on average, while strong priors produced more variable results. For every external trial, at least one BPSM outperformed all PSMs, and seven of eight matched the long-term Kaplan-Meier curve (ΔAUC<1 month). AIC, BIC and DIC on the Bayesian fit poorly predicted accuracy; goodness-of-fit on the external trial provided a useful but imperfect guide to distribution selection.
CONCLUSIONS: BPSMs prevented the systematic overestimation observed with standard PSMs and should be explored in HTA submissions. Distribution selection should always be validated against clinical expert opinion to ensure long-term plausibility.