Causal Machine Learning Analysis of Radiation-Induced Leukemia and Solid Tumor Incidence in Japanese Atomic Bomb Survivors

Uncertainty in low-dose ionizing radiation-induced health risks stems from several factors. The complex biological pathways leading to diseases like cancer are not fully understood, making it difficult to distinguish the contribution of radiation, particularly at low doses which induce only small perturbations to background disease risks. Additionally, traditional dose-response models, such as the Linear No-Threshold formalism and competing threshold or hormesis models, impose rigid assumptions on dose response shapes, causing controversy and increasing model selection uncertainty. Furthermore, these modeling strategies operate on the level of correlations/associations, and are not designed to directly address the ultimate goal of radiation epidemiology - assessing causal links between radiation and disease. A promising and rapidly-developing approach for addressing some of these challenges is causal machine learning (CML), such as double/debiased machine learning (DML), which is designed to model causal effects in multi-dimensional data sets. Our study employs DML to elucidate the causal impacts of radiation exposure on the incidence of leukemia, all solid tumors, and stomach tumors among Japanese atomic bomb survivors. Its goal was not to produce a definitive re-analysis of these data sets, but to provide a useful example of implementing CML in radiation epidemiology, which can advance the field by supplementing traditional modeling approaches. The results revealed robust positive causal effects of radiation for all three tumor types, especially for leukemia and stomach tumors. The effect magnitudes, and uncertainties, were not dramatically different at low doses than at higher doses. The influences of age at exposure, attained age, sex and other covariates on the causal effects of radiation were assessed using Shapley Additive Explanations (SHAP) values. We believe that this analysis, based on a flexible machine learning framework with a causal inference motivation and without strict dose response assumptions, provides an important contribution to radiation epidemiology.