Presentation (CTI)

OBJECTIVES: Estimates of future survival play a critical role in health technology assessments (HTA). While standard parametric models are commonly used for survival extrapolation, flexible models have shown improved within-sample fit but do not deliver more accurate future projections. This study seeks to address whether survival extrapolations should focus on better fit to the available data or minimizing future uncertainties when survival data are immature.
METHODS: Survival curves were digitalized and corresponding pseudo-patient-level data, including censoring were reconstructed for overall survival (OS) and progression-free survival (PFS) data from 10 clinical trials with mature survivals. To simulate immature data, the datasets were artificially censored at 60-70%, 50%, 30%, and 20% event thresholds. Extrapolations were performed using five standard parametric functions (generalized gamma, Weibull, exponential, log-normal, log-logistic) for the complete and artificially censored datasets. Extrapolated life-years (LY) and progression-free time (PFT) were compared to the Kaplan-Meier (KM) estimates to quantify future accuracy using a partitioned survival model framework.
RESULTS: Most parametric functions produced similar projections (within 1.5%) compared to the KM estimates with the complete dataset. However, with increasing event thresholds censorship, uncertainty for LY and PFT estimates increased from ± 10% at 60%-70% threshold up to -70% to +40% at 20-30% threshold. Moreover, the performance of future projection for each function varied with log-normal and exponential functions showing ±6% uncertainty overall, while generalized gamma exhibited significant instability with increasing censoring thresholds.
CONCLUSIONS: These preliminary findings, based on a limited sample of studies, highlight a critical challenge in HTA decision-making: current guidelines prioritize fit to available KM data, potentially overlooking uncertainties in future projections when data are immature. To improve decision-making, an alternative approach emphasizing the choice of functions to produce better long-term predictive accuracy over short-term fit may be warranted.