Causal Machine Learning Analysis of All-Cause Mortality in Japanese Atomic-Bomb Survivors

The health consequences of ionizing radiation have long been studied, yet significant uncertainties remain, particularly at low doses. In particular, traditional dose-response models such as linear, linear-quadratic, threshold, or hormesis models, all impose specific assumptions about low-dose effects. In addition, while the goal of radiation epidemiological studies is ideally to uncover causal relationships between dose and health effects, most conventional data analysis techniques can only establish associations rather than causation. These limitations highlight the need for new analysis methodologies that can eliminate the need for a priori dose-response assumptions and can provide causal inferences more directly based on observational data. Causal Machine Learning (CML) is a new approach designed to uncover how changes in one variable directly influence another, and with these motivations, a CML approach was, for the first time, implemented here to analyze radiation epidemiological data – in this case all-cause mortality data from Japanese A-bomb survivors. Compared to more traditional parametric approaches for analyzing radiation epidemiological data such as Poisson regression, CML makes no a priori assumptions about dose-effect response shapes ( e.g., linearity or thresholds). Extensive validation and refutation tests indicated that the proposed CML methodology is robust and is not overly sensitive to unmeasured confounding and noise. At moderate to high radiation doses, the CML analysis supports a causal increase in mortality with radiation exposure, with a statistically significant positive average treatment effect (p = 0.014). By contrast, no statistically significant causal increase in all-cause mortality was detected at doses below 0.05 Gy (50 mGy). These conclusions were drawn after adjusting for all available key covariates including attained age, age at exposure, and sex. We emphasize that this CML-based approach is not designed to validate or disprove any particular dose-response model. Rather this approach represents a new potentially complementary approach that does not rely on a priori functional form assumptions.