Dose-Response after Low-dose Ionizing Radiation: Evidence from Life Span Study with Data-driven Deep Neural Network Model

Accurately evaluating the disease risks after low-dose ionizing radiation (IR) exposure are crucial for protecting public health, setting safety standards, and advancing research in radiation safety. However, while much is known about the disease risks of high-dose irradiation, risk estimates at low dose remains controversial. To date, five different parametric models (supra-linear, linear no threshold, threshold, quadratic, and hormesis) for low doses have been studied in the literature. Different dose-response models may lead to inconsistent or even conflicting results. In this manuscript, we introduce a data-driven deep neural network (DNN) model designed to evaluate dose-response models at low doses using Life Span Study (LSS) data. DNNs possess the capability to approximate any continuous function with an adequate number of nodes in the hidden layers. Being data-driven, they circumvent the challenges associated with misspecification inherent in parametric models. Our simulation study highlights the effectiveness of DNNs as a valuable tool for precisely identifying dose-response models from available data. New findings from the LSS study provide robust support for a linear quadratic (LQ) dose-response model at low doses. While the linear no threshold (LNT) model tends to overestimate disease risk at very low doses and underestimate health risk at relatively high doses, it remains a reasonable approximation for the LQ model, given the minor impact of the quadratic term at low doses. Our demonstration underscores the power of DNNs in facilitating comprehensive investigations into dose-response associations.