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

Abstract

Chronic radiation sickness is a deterministic radiation health effect observed among the Mayak Production Association workers in Russia. In this study, unsupervised neural networks were used to cluster hematological measurements in a subset (n = 88) of the Mayak Production Association population while excluding from the analysis the radiation dose and the historical clinical diagnosis. Clusters of observations that had lower average leukocyte and thrombocyte counts were labeled "affected" and those having higher average blood cell counts were labeled "unaffected." The class (cluster) membership for each individual was used subsequently as a dependent variable in a classification tree model in order to identify significant features of the underlying classification model. After re-classification of cases using this method, the results showed a better data separation between the blood cell counts for affected vs. unaffected groups compared to those based on historical classification, and a greater difference between group means for differential blood counts was observed than for the historical diagnosis. The reclassification of diagnostic groups changed the group mean radiation doses. The geometric means (and 95% CL) of cumulative radiation dose equivalent from external exposures, based on the historical diagnosis, are 0.31 (0.0035, 3.4) vs. 1.7 (0.0007, 18) Sv. After clustering and classification tree analyses, the group geometric means were 0.78 (0.0014, 8.6) vs. 1.5 (0.0007, 17) and 0.82 (0.0013, 9.0) vs. 1.4 (0.0008, 16) Sv, using (respectively) whole blood cell counts or differential counts as the independent variables. The approach presented here is useful as a diagnostic aid for both retrospective analyses and in the event of future radiation accidents.