Abstract
Descendants of Semipalatinsk Nuclear Test Site-exposed individuals have a documented multigenerational health burden, yet parental dose-integrated predictive modelling of descendant outcomes remains absent. This study evaluates the feasibility of ICD chapter-level outcome group classification in registry-derived descendant cohorts using machine learning, with both cohorts restricted to outcome-positive individuals. Two cohorts were derived from the State Scientific Automated Medical Registry: a second-generation cohort (n = 7898) and a three-generation pilot cohort (n = 2841). Logistic Regression, Random Forest, XGBoost, and LightGBM were evaluated via stratified five-fold cross-validation using eight parental dose and demographic predictors; no descendant-level clinical measurements were included. In the second-generation cohort, AUC-ROC reached 0.819 for cardiovascular outcomes, 0.776 for neoplasms, and 0.835 for the composite endpoint. A unified-outcome sensitivity analysis incorporating death-certificate and protocol records increased second-generation positive counts by 22.3–24.7%, while discrimination remained broadly stable. Birth year was the dominant predictor, reflecting biological ageing, ascertainment differences, and exposure-era confounding; dose-only models retained AUC values of 0.619–0.761, indicating that the recorded parental dose variables contained above-chance predictive information within this registry cohort, without establishing a biological, causal, or heritable intergenerational effect. Third-generation AUC values of 0.901–0.918 were stable across the conducted resampling checks but were most plausibly driven by structural birth-year confounding. These results establish a reproducible performance baseline for this nuclear legacy registry setting and identify healthy cohort inclusion, age-standardised ascertainment, and individual biological markers as key prerequisites for advancing this research.