EX1 How to Select the Appropriate Survival Extrapolation Method? a Proposed Framework

Abstract

For HTA, apart from standard parametric models (SPMs), other extrapolation approaches are applied such as piecewise, splines, mixture cure and non-mixture cure, parametric-mixture, and fractional polynomial models. The objective is to validate a survival extrapolation decision framework suggesting cure models for curative therapies, parametric-mixture models in case of a subgroup of long-term survivors, piecewise, splines and fractional polynomials in case of fluctuating hazards, and SPMs for other data. All survival extrapolation approaches were fitted on immature data of three trials; in melanoma the framework recommends a cure model as cure seems clinically plausible, in breast cancer (BC) a parametric-mixture as a subgroup of long-term survivors is expected, and in multiple myeloma (MM) an SPM as no cure, long-term survivors or fluctuating hazards are expected. The statistical fit of the models was assessed using LOOIC. The extrapolation on immature data was compared to mature trial data in terms of ΔMean-Absolute-Deviation (ΔMAD) in months for all arms in the trial. Lower LOOIC and ΔMAD imply better predictions. Restricted mean survival time (RMST) of the mature data was determined to assess the relative size of ΔMAD. The best fitting approach on the melanoma dataset (RMST: 67.88 months) was the log-logistic parametric-mixture (LOOIC=2,870, ΔMAD=19.59). The best fitting cure model was the lognormal (LOOIC=3,249, ΔMAD=10.89). On the BC dataset (RMST: 123.42 months), the lognormal parametric-mixture fitted best (LOOIC=1,782, ΔMAD=9.86) and had one of the lowest ΔMAD. On the MM dataset (RMST: 92.55 months), the Weibull parametric-mixture fitted best (LOOIC=1,944, ΔMAD=11.24), and the best fitting SPM was the exponential (LOOIC=1,988, ΔMAD=7.21). This study shows that the model recommended by the framework generally leads to better long-term extrapolations compared to the statistically best fitting model. Therefore, additional to historical data and long-term comparative effectiveness expectations, this framework can be used for model selection.