Machine Learning-Driven Dose Estimation In Biodosimetry: An Alternative to Calibration Curves In Low Dose-Protracted Exposure Scenarios
Abstract
Purpose
To determine whether a supervised machine learning (SML) approach to low-dose biodosimetry, trained on radiation-induced chromosomal translocations of astronauts and individual radiation response covariates, could be an alternative method for dose estimation.
Methods
Two astronaut study groups were established by laboratory of analysis, and compared (35 American astronaut samples; National Aeronautics and Space Administration (NASA) and 10 Canadian and European astronaut samples; Health Canada (HC)). Dose response curves were generated by enumerating translocations in irradiated (X-ray or γ-ray doses up to 2 Gy) whole blood drawn from astronauts prior to long missions (> 3 months) to the International Space Station. Blood samples were also collected shortly after and between six to eighteen months after flight. Translocations were measured using whole chromosomes fluorescence in situ hybridization (FISH) in metaphase (HC and NASA) or premature condensed interphase chromosomes (NASA). GradientBoosting was chosen after training on several SML algorithms on endpoint yields from samples with complete datasets and covariates including age and sex. Photon equivalent doses were estimated by referencing the first post-flight translocation yields to individual astronaut calibration curves and the SML approach. Doses were compared to whole-body effective doses, generated according to the NSCR-2012 formalism.
Results
Statistical analysis of the translocations showed significant increases of translocations in over 75% of endpoints measured post-flight. Dose estimates from calibration curve fits were 50(±24) mSv (HC) and 95(±62) mSv (NASA) while the SML algorithm average dose estimates were 59(±43) mSv (HC) and 26(±39) mSv (NASA). SML performed more robustly when good quality CCs were not obtained.
Conclusion
This study confirms exposure to ionizing radiation from long-duration ISS missions typically results in significant increases of translocations in the lymphocytes of astronauts. These responses demonstrate the feasibility of a supervised machine learning approach, especially when pre-exposure data is suboptimal or missing.