Causal re-assessment of the spatial association between normal tissue dose and overall-survival in lung radiotherapy.

Abstract

Voxel-based analysis (VBA) aims to analyse spatial dose distributions to identify anatomical (sub)-regions associated with outcomes. Multivariable models used in this approach are not geared to infer causality; to achieve causality, this work aims to utilise directed acyclic graphs (DAGs) to produce minimal causal adjustment sets for four publications. One VBA paper is reanalysed as an exemplar for how causal principles should be applied analytically. A DAG was developed to model the causal effect of excess dose to the region of significance identified in the base of the heart on OS. Covariates controlled for in the original papers were adjusted on the DAG, showing biasing pathways remained open, prompting minimal adjustment sets to be devised.1100 NSCLC patients from a previous VBA analysis, treated with 55 Gy in 20 fractions were reanalysed. Multivariable Cox regression analysis compared the original analysis with the minimal causal adjustment set. Hazard ratios (HR) and 95% confidence intervals were calculated. Common covariates identified across all minimal adjustment sets were: staging, radiotherapy modality, tumour location and prescribed dose. Lung dose, ejection fraction, and adverse events introduced potential bias in the original analyses. Our exemplar reanalysis showed excess dose to the heart base is time dependent, increasing risk of mortality ≤ 18 months follow up - HR: 1.133 (1.065-1.207), after which no effect was observed - HR: 0.948 (0.873-1.030). This work reasons that causal theoretical principles should be used to minimise potential bias in studies utilising VBA for research into RT adverse events and survival.