Classification of chronic radiation sickness cases using neural networks and classification trees

Area of Science:

• Hematology
• Radiation Health Effects
• Biostatistics

Background:

• Chronic radiation sickness is a deterministic effect observed in Mayak Production Association workers.
• Traditional clinical diagnoses may not fully capture the nuances of radiation exposure effects.

Purpose of the Study:

• To apply unsupervised neural networks for clustering hematological data.
• To re-classify Mayak workers based on blood cell counts, excluding radiation dose and historical diagnosis.
• To identify significant features differentiating affected and unaffected groups.

Main Methods:

• Unsupervised neural networks for clustering hematological measurements (leukocytes, thrombocytes).
• Classification tree models to identify differentiating features based on cluster membership.
• Exclusion of radiation dose and historical diagnosis during initial clustering.

Main Results:

• Clusters with lower leukocyte and thrombocyte counts were labeled 'affected'.
• Re-classification showed better data separation and greater differences in differential blood counts compared to historical diagnosis.
• Re-classification altered mean radiation dose estimates for the groups.

Conclusions:

• The neural network and classification tree approach offers a refined method for analyzing radiation exposure effects.
• This method serves as a valuable diagnostic aid for retrospective studies and future radiation incidents.
• Improved classification of affected vs. unaffected groups based on hematological data.