PCN20 ACCURATE PREDICTIONS OF LIFE YEAR GAINS FOR IMMUNO-ONCOLOGY THERAPIES IN THE LONG TERM? AN ANALYSIS BASED ON PUBLISHED CHECKMATE 057 NIVOLUMAB DATA

Abstract

Immuno-oncology (I-O) therapies are associated with complex hazard functions due to delayed treatment effects and long-term survivors. Therefore, using best statistical fit to determine the extrapolation of short-term trial data at the time of an HTA submission may lead to inaccurate predictions of long-term survival. This research aimed to retrospectively analyse the accuracy of different extrapolations performed at early data cuts in predicting realised long-term life years (LYs) based on the CheckMate 057 study of nivolumab (NCT01673867). Over a 51.8-month time horizon (corresponding to the longest available duration of observed Kaplan-Meier [KM] data), cumulative LYs were estimated for standard parametric models and spline models (1–4 knots) fitted to digitised published overall survival data for nivolumab from successive interim data cuts of the CheckMate 057 trial in advanced non-squamous non-small-cell lung cancer (minimum follow-ups of 13.2, 18.0 and 24.2 months). Estimated LYs at each interim data cut were compared to LYs calculated from published long-term KM data observed over 51.8 months to determine which interim models provided the most accurate predictions. Estimates from the spline 1 knot (13.2 months minimum follow-up) and the log-logistic model (18.0 and 24.2 months minimum follow-ups) most accurately predicted the LYs accumulated over 51.8 months from observed KM data. Notably, these contrast with the best statistically fitting models at the interim data cuts. This study provides empirical evidence that the model of best statistical fit for short-term survival data may not provide the most accurate estimates of LYs for I-O therapies over the long-term, which may lead to inaccurate cost-effectiveness estimates of treatments. Further research is required to determine whether there are learnings that are generalisable across I-O therapies and indications to support the use of particular model choices, which allow for long-term survivors, at early data cuts.