Projecting overall survival data for health-economic models in oncology: Do maturity levels impact uncertainty?

Abstract

e18350 Background: A lifetime horizon is recommended for health-economic evaluation of anticancer drugs. If overall survival (OS) data is immature, extrapolation of the Kaplan-Meier (KM) curve using distributions is done to obtain long-term data. Depending on OS maturity, the distribution chosen may impact estimation of life expectancy (LE) and of life years gained (LYG) between treatments. This study aimed to estimate the error (on LE and LYG) induced by the choice of extrapolation distributions, for 2 levels of OS maturity (30% and 50%), as compared to LE and LYG at full maturity. Methods: Fifteen phase 3 trials published between 2013 and 2017 containing OS KM curves were selected if maturity > 70% (Full). To test 2 maturity levels, each KM curve was truncated at 30% and 50%. A 3-step process was performed: 1) KM was digitalized, 2) individual patient-data was reproduced using the Guyot algorithm, and 3) 5 parametric survival distributions were fit using the R-Survival package. For each study, the process was done for each treatment arm and each of the 3 maturity levels on the same time horizon, equal to the maximum follow-up of the study. For each curve, the best distribution was chosen by a board of 2 oncologists and 2 health-economists, based on visual inspection, Akaike/Bayesian Information criteria and external validity. Results: Based on the board review of the 90 KM curves, main chosen distributions were Weibull (33%), Log-logistic (32%) and Log-normal (27%). As compared to LE at full maturity, LE was overestimated in 23% and 40% at 30% and 50% maturity, respectively. Mean absolute error was 0.88 at 30% maturity, and decreased to 2.12 at 50% maturity, i.e. estimation was 2.4 times better from 30% to 50% maturity. On average, at 30% maturity versus full, mean percentage of error in LYG was 126.4% and 62.4% at 50% maturity versus full. Conclusions: The use of immature OS data increases the risk of error when projecting long-term LE. Even marginal gains in OS maturity can be translated in more accurate estimations and guide health-economist in developing models.