Machine learning approach for quantitative biodosimetry of partial-body or total-body radiation exposures by combining radiation-responsive biomarkers

During a large-scale radiological event such as an improvised nuclear device detonation, many survivors will be shielded from radiation by environmental objects, and experience only partial-body irradiation (PBI), which has different consequences, compared with total-body irradiation (TBI). In this study, we tested the hypothesis that applying machine learning to a combination of radiation-responsive biomarkers (ACTN1, DDB2, FDXR) and B and T cell counts will quantify and distinguish between PBI and TBI exposures. Adult C57BL/6 mice of both sexes were exposed to 0, 2.0-2.5 or 5.0 Gy of half-body PBI or TBI. The random forest (RF) algorithm trained on 1/2 of the data reconstructed the radiation dose on the remaining testing portion of the data with mean absolute error of 0.749 Gy and reconstructed the product of dose and exposure status (defined as 1.0 x Dose for TBI and 0.5 x Dose for PBI) with MAE of 0.472 Gy. Among irradiated samples, PBI could be distinguished from TBI: ROC curve AUC = 0.944 (95% CI: 0.844-1.0). Mouse sex did not significantly affect dose reconstruction. These results support the hypothesis that combinations of protein biomarkers and blood cell counts can complement existing methods for biodosimetry of PBI and TBI exposures.

Automated summary

In this study, machine learning was applied to a combination of radiation-responsive biomarkers and B and T cell counts to quantify and distinguish between partial-body irradiation (PBI) and total-body irradiation (TBI) in mice. The results showed that the random forest algorithm could accurately reconstruct the radiation dose and differentiate between PBI and TBI. These findings suggest that protein biomarkers and blood cell counts can complement existing methods for biodosimetry of PBI and TBI exposures.