Calibration Algorithms for Estimating the Number of Radiogenic Deaths in a Cohort

Abstract

Background: Development of algorithms for estimating the number of radiogenic deaths in a cohort that happened or may happen before an arbitrary moment in time. Algorithms should be based on continuous lifelong risk models and minimize the errors of estimates based on deformable non-radiation models ("seed" models) reflecting the general patterns of Russian mortality, but with the possibility of calibrating according to selected data. Materials and methods: For the study, the databases of the register of liquidators of the consequences of the Chernobyl accident – employees of the enterprises of the Ministry of Atomic Energy of Russia were used. Persons with a non-zero dose of ionizing radiation received as a result of the accident and who were registered in the register until 2018 were selected from the register. There are 7324 men and 655 women in this sample. The number of deaths from the beginning of the observation period until 1996, 2006, 2016 from all causes, as well as from solid cancers and leukemias, was used as a reference data for the calibration. Radiation risk was calculated using UNSCEAR 2006 models. Calibration was performed using the Levenberg – Marquardt method. Results: A general formula was obtained for assessing the excess lifetime risk of death, which is the basis of the developed algorithms. The "seed" models of the risk of death from all causes and from solid cancers and leukemia’s were obtained. In accordance with the purpose of the work, two algorithms have been developed. The first algorithm is intended for direct (without using radiation risk models) estimation of the number of radiation-induced deaths in a cohort. The second uses radiation risk models. The performance of the algorithms was tested using the industry register of liquidators of the consequences of the Chernobyl accident. In particular, a 30 % calculated increase in the estimated number of deaths from radiation-induced solid cancers was recorded using calibrated models. Conclusions: The proposed calculation algorithms for calibrated models provide reliable (if the radiation risk models are considered sufficiently adequate) estimates of the number of Chernobyl deaths for points in time preceding the current one and, presumably, more accurate predictive estimates compared to the calculation algorithm for uncalibrated models. The low sensitivity of the lifelong risk model to variations in the parameters of the background models and the total mortality model makes it possible to apply models calibrated according to the liquidator register to other closely related populations (NPP workers, etc.).