Prediction Error in Quality-Adjusted Life Years in Economic Evaluations of Immune Checkpoint Inhibitors: A Comparison Based on Projected and Observed Updated Survival.

Abstract

Survival data from pivotal clinical trials are critical for estimating quality-adjusted life years (QALYs). However, immature survival data require extrapolation beyond observed follow-up to project outcomes, introducing potential prediction error into QALY estimates used in economic evaluations. Immune checkpoint inhibitors (ICIs) present unique extrapolation challenges due to delayed responses and extended survival benefits. We therefore quantify the QALY prediction error of early extrapolations by benchmarking them against updated follow-up at a common time horizon. Using reconstructed individual patient data derived from published Kaplan-Meier curves of pivotal trials, this study assessed the accuracy of early survival extrapolations for ICIs approved in China by comparing early projections with QALYs obtained from updated data at the same horizon. A partitioned-survival framework using overall survival (OS) and progression-free survival (PFS) informed state occupancy, and QALYs were obtained via restricted mean survival time (RMST) integration of health-state utilities at the target horizon. Statistical analyses evaluated bias, precision, and agreement between extrapolated and updated QALY estimates. Linear regression and sensitivity analyses assessed the impact of target extrapolation horizon (T) on prediction error. In total, 14 randomized controlled trials (4839 patients) were included for analysis. The mean deviation between extrapolated and observed QALYs was - 0.01 (95% CI - 0.03 to 0.01), with a mean absolute error (MAE) of 0.03 (95% CI 0.01 to 0.04). Strong agreement existed between extrapolated and updated QALYs (Spearman's ρ = 0.98, 95% CI 0.94 to 0.99, P < 0.001). Consistently, OS extrapolations showed minimal deviation (MAE: 0.56 months), while PFS tended to be underestimated (MAE: 1.91 months). Moreover, predictive error increased significantly with longer extrapolation periods for QALY (MAE increase: 0.011 QALY/year, P = 0.004) and OS (MAE increase: 0.342 months/year, P = 0.010). Near-horizon QALY prediction error was modest on average but increased with longer target extrapolation horizons. These findings support transparent reporting of extrapolation uncertainty and suggest that structured evidence reassessment may be particularly valuable as longer follow-up becomes available.