Application of Random Forest Algorithm for Dose Reconstruction Based On Classification of Clean-Up Workers of the ChNPP Accident
K Chizhov1*, V Kryuchkov2, V Chumak3, C Lee1, V Drozdovitch1, E Cahoon1, V Vij1, R Zhang1, I Golovanov2, E Bakhanova3, N Gudzenko3,Bazyka D3, M Little1, (1) National Cancer Institute, Rockville, MD, (2) State Research Center Burnasyan Federal Medical Biophysical Center of Federal Medical Biological Agency, Moscow, Russia, (3) National Research Centre for Radiation Medicine, Kyiv, Ukraine
Presentations
MO-IePD-TRACK 6-1 (Monday, 7/26/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Purpose: As a result of the accident at Chernobyl Nuclear Power Plant a heavy contamination of the environment occurred. Thousands of workers took part in cleanup work, and only few percent of them have sufficient dosimetry. Thus, methods are required for dose reconstruction for large groups of liquidators based on a limited set of data.
Methods: A dosimetry method was developed based on random forest algorithm. It was applied to a cohort of 11105 liquidators from Clinical and epidemiological registry of Ukraine. Neural network was trained (70%) and tested (30%) on database of 1807 liquidators, whose doses were reconstructed according to detailed questionnaires using the Realistic Analytical Dose Reconstruction with Uncertainty Estimation (RADRUE) method. The developed algorithm was based on the date of the mission and duration, occupation, place of work and residence, affiliation to the organization. Dose reconstruction for study cohort was made in 2 steps: reconstruction of the specific group (contingent) of the liquidator and then dose assessment within the contingent. Rules for division into continents described in RADRUE method.
Results: Liquidators from study cohort were divided into 9 contingents. The contingents were also divided by year from 1986 to 1990. Based on RADRUE database the dose distribution for each contingent any year was calculated. The developed method determines with an accuracy of 91.5% whether the liquidator was a military or civilian. For other contingents, the correctness of the method was from 73.7% to 99.8%.
Conclusion: Developed algorithm applicable for contingent and dose range reconstruction for liquidator if the training dataset contains all possible variants from the studied dataset. Results of the study and the cohort will be used as a resource to investigate novel hypotheses about the radiation-related risk of cancer and benign outcomes by combining historical dose records, epidemiological data and medical records.
ePosters
Keywords
Taxonomy
IM/TH- Mathematical/Statistical Foundational Skills: Machine Learning
Contact Email