Causal analysis of radiation dose to the base of the heart and overall survival in lung cancer using targeted maximum likelihood estimation
B Jiang, M Summers, E V Osorio, K Banfill, J Hartfold, C Faivre-Finn, A McwilliamAbstract
Background
Radical radiotherapy is delivered with curative intent in approximately 40% of patients with lung cancer. The RTOG 0617 trial and subsequent studies demonstrated that higher cardiac radiation exposure is associated with increased mortality. As clinical equipoise precludes randomisation to different cardiac dose limits, estimation of the causal effect of cardiac radiation dose on survival must rely on observational data. Traditional analyses have relied on multivariable regression models incorporating heart dose parameters and cardiac risk factors, but failure to consider causal structure may induce confounding bias. Targeted maximum likelihood estimation (TMLE) integrates machine learning to estimate outcome and treatment assignment models, approximating a randomised comparison within observational data and providing reliable effect estimates even if one model is misspecified.
Purpose
Previous work has identified radiation dose to the base of the heart as being associated with increased mortality. However, the causal effect of radiation dose to the base of the heart on overall survival remains unclear. We therefore aimed to estimate this effect using a target trial emulation framework with TMLE.
Methods
The study included 1,051 patients with lung cancer treated with radiotherapy at a tertiary cancer centre between 2010 and 2013, with time zero defined as radiotherapy initiation. Radiation dose to the base of the heart was categorized as low (0–11.26 Gy), medium (11.26–23.34 Gy), or high (23.34–48.85 Gy). Overall survival was defined from radiotherapy initiation to death or censoring at last follow-up. The causal estimand was defined as time-specific differences in survival probability between dose groups. A directed acyclic graph (DAG) identified the minimal adjustment set. TMLE was applied to compare overall survival across dose groups over a 2-year follow-up.
Results
After dose categorization, 367 patients were included per group, with approximately 300 retained per group in complete-case analyses. The DAG identified age, tumor volume, tumor location, T stage, M stage, and radiotherapy modality as the minimal adjustment set. No meaningful survival difference was observed between the low- and medium-dose groups. In the high- versus low-dose comparison, survival differences emerged early and became consistent after approximately 5 months, peaking during mid follow-up and remaining evident through the end of follow-up. In the high- versus medium-dose comparison, survival differences showed transient mid follow-up significance and did not become consistently significant until approximately 15 months.
Conclusion
Under causal assumptions within regions of sufficient overlap, higher radiation dose to the base of the heart was associated with distinct time-varying survival differences. Further studies should explore dose–response patterns across cardiac substructures within a causal analysis framework.High–low survival difference (TMLE) High–medium survival difference (TMLE)