A machine learning method for improving the accuracy of radiation biodosimetry by combining data from the dicentric chromosomes and micronucleus assays

A large-scale malicious or accidental radiological event can expose vast numbers of people to ionizing radiation. The dicentric chromosome (DCA) and cytokinesis-block micronucleus (CBMN) assays are well-established biodosimetry methods for estimating individual absorbed doses after radiation exposure. Here we used machine learning (ML) to test the hypothesis that combining automated DCA and CBMN assays will improve dose reconstruction accuracy, compared with using either cytogenetic assay alone. We analyzed 1349 blood sample aliquots from 155 donors of different ages (3–69 years) and sexes (49.1% males), ex vivo irradiated with 0–8 Gy at dose rates from 0.08 Gy/day to ≥ 600 Gy/s. We compared the performances of several state-of-the-art ensemble ML methods and found that random forest generated the best results, with R(2) for actual vs. reconstructed doses on a testing data subset = 0.845, and mean absolute error = 0.628 Gy. The most important predictor variables were CBMN and DCA frequencies, and age. Removing CBMN or DCA data from the model significantly increased squared errors on testing data (p-values 3.4 × 10(–8) and 1.1 × 10(–6), respectively). These findings demonstrate the promising potential of combining CBMN and DCA assay data to reconstruct radiation doses in realistic scenarios of heterogeneous populations exposed to a mass-casualty radiological event.