Accurately estimating radiation risk is crucial for protecting public health, establishing safety standards, and advancing research in radiation safety. Historically, epidemiological studies have relied on parametric risk models, which are simple and interpretable but can yield risk estimates that are highly dependent on model specifications. Deep neural networks (DNNs), in contrast, make fewer assumptions and have driven significant advances across many fields. However, the full potential, limitations, and implications of DNNs for radiation risk assessment remain largely unexplored. In this study, we evaluated the feasibility and practicality of using DNNs for radiation risk assessment, using the RERF Life Span Study (LSS) cohort as our dataset. We found that DNNs offer greater flexibility and model independence, achieving slightly better or comparable performance in predicting tumor incidence rates across predefined metrics. Interestingly, excess relative risk (ERR) estimates from DNNs differed significantly from those of the parametric model, despite similar predictive accuracy for incidence rates. In the parametric models, ERR estimation can be influenced by the specified functional forms for both baseline risk and age-related effect modification. To investigate the source of this discrepancy, we applied SHAP values, which indicated that the DNN identifies radiation dose as the primary contributor to ERR, offering an alternative perspective. Although DNNs have limitations, such as reduced interpretability and challenges in constructing confidence intervals, they may serve as a complementary tool to well-specified parametric models for exploring complex nonlinear patterns and interactions.
Liu et al. (Mon,) studied this question.