Presentation (CTI)

OBJECTIVES: Parametric extrapolation of clinical-trial survival data is standard in health technology assessment, but it relies on distributional assumptions that are sometimes difficult to validate, particularly given the lack of long-term data. Random Survival Forests (RSF) have shown promise for long-term extrapolation beyond parametric methods. This study assessed whether RSF could be used as a reliable extrapolation method, using data simulated from a single source to emulate trial-like and real-world-like populations.
METHODS: The National Lung Screening Trial dataset (1,749 lung-cancer patients) served as a single-source framework. Divergence was induced using a Cox-based prognostic score reflecting age, cancer stage and comorbidities: the real-world-like cohort comprised patients with less favorable prognostic profiles, and trial-like follow-up was truncated at 24 months. Divergence was controlled by a parameter δ (1.0-2.5). Three approaches were compared against observed Kaplan-Meier survival: parametric models fitted to the truncated trial-like data, a RSF combining the two cohorts, and a RSF using only the trial-like population. RSF models used 21 selected predictors; performance over a 7-year horizon was assessed using mean absolute error (MAE).
RESULTS: The combined RSF reconstructed long-term survival markedly better than parametric extrapolation (mean MAE 0.016 versus 0.036 for the best parametric model; Gompertz worst at 0.19), an advantage holding across all divergence levels with MAE remaining low, indicating stable performance. The trial-like-only RSF performed poorly (MAE 0.18), unable to project beyond the last observed event after truncation: the gain stemmed from incorporating long-term data, not from intrinsic extrapolation ability.
CONCLUSIONS: Within a single-source framework, combining truncated trial-like and long-term real-world-like data in an RSF provided more accurate and robust survival reconstruction than parametric extrapolation, with improvements driven by incorporation of real-world-like data rather than by the algorithm itself. Future research should evaluate this approach using multiple data sources combining clinical-trial and real-world evidence.