PCN413 A BAYESIAN MIXTURE CURE MODEL USING INFORMATIVE PRIORS

Abstract

The use of mixture cure models (MCMs) is becoming more relevant within health technology assessment (HTA) of curative cancer treatments to extrapolate overall survival (OS). At HTA submission, survival data may not be mature enough to reliably estimate cure and extrapolate survival for both treatment arms. To improve the accuracy in MCMs, progression-free survival (PFS) cure rates or published mature OS data can be used as priors to inform OS. In this study, a Bayesian MCM with priors is introduced to predict immature OS rates in previously untreated advanced melanoma. The MCM with priors was tested on immature data from an RCT assessing nivolumab versus dacarbazine. The PFS curve in this trial showed a cure rate of 40% for nivolumab. Mature trial data assessing ipilimumab+dacarbazine versus dacarbazine showed a cure rate of 10% for dacarbazine. The cure rates were incorporated as prior information using a logit model. of the MCM with priors were compared to the MCM without priors and standard Weibull distribution. The predicted cure rates and survival extrapolations for the three approaches were validated with mature OS data from the long-term follow-up study of nivolumab versus dacarbazine. The MCM with priors estimated cure rates of 37.48% and 9.75% for nivolumab and dacarbazine, respectively compared to 19.99% and 0.65% without priors. The mean OS was 8.1 and 2.6 years with priors, 4.4 and 1.1 years without priors, and 4.3 and 1.0 years with standard Weibull for nivolumab and dacarbazine, respectively. The validation data closest resembled the survival predictions of the MCM with priors, followed by Weibull and MCM without priors. The use of informative prior distributions produced the most accurate results compared to standard Weibull extrapolation and MCM without priors. Prior derivation based on literature and expert opinion can improve the fit of mixture cure models considerably.