1. Data
The primary source is the SSAMR registry (State Scientific Automated Medical Registry), a relational database of 37 tables covering demographic records, exposure history, medical outcomes, and family relationships. After systematic filtering — requiring at least one registered outcome and dose data for at least one parent — the core second-generation cohort was formed. Average data missingness dropped from 60.2% in the raw registry to 14.2% in the analytical cohort.
7 898
14.18%
average missingness
1966
median birth year
51.8%
male sex
8 predictors
Dose-related (6)
father_aggregated_dose
father's agg. dose, mGy
mother_aggregated_dose
mother's agg. dose, mGy
max_parent_dose
maximum parental dose
mean_parent_dose
mean parental dose
father_dose_available
father's dose availability indicator
mother_dose_available
mother's dose availability indicator
Demographic (2)
birth_year
birth year (age proxy)
sex_binary
sex (1 = male, 0 = female)
3 target variables
Chapter I
16.5%
Cardiovascular diseases · n = 1,304
Chapter C
6.1%
Neoplasms · n = 482
Composite C/I
22.0%
Combined endpoint · n = 1,739
Secondary endpoint; not interpreted as a single clinical entity
2. Algorithms
Three model families were trained in parallel on each target variable — from a linear baseline to ensemble methods. All performance estimates were obtained via stratified 5-fold cross-validation: predictions on each fold were produced by a model that had not seen those observations during training.
Logistic Regression
A linear baseline with L2 regularization (ridge, C = 1.0). Establishes a linear separating surface; interpretable and serves as a lower bound for comparison.
Random Forest
A non-parametric ensemble of 300 decision trees (max depth 6). Captures non-linear interactions by averaging over independent bootstrap samples.
XGBoost
Gradient-boosted trees (300 trees, lr = 0.05, depth 4). Sequentially approximates residuals; the primary non-linear model with the highest AUC.
3. Model Results
The primary metric is AUC-ROC, which evaluates ranking ability independent of the classification threshold. Average Precision complements the assessment for imbalanced classes (especially Chapter C, with a 6.1% event rate). Brier Score measures the calibration of predicted probabilities.
AUC-ROC by target variable and model
| Target | Model | AUC-ROC | Avg Prec | Brier |
|---|
4. Feature Importance
About the method
SHAP (SHapley Additive exPlanations) is a prediction-explanation method based on cooperative game theory. For each feature, the Mean |SHAP value| is computed — the average absolute change in the model's prediction attributed to that feature. A feature with a larger value shifts the prediction more strongly relative to the baseline.
Birth year — the dominant feature
birth_year ranks first by a wide margin (Mean |SHAP| ≈ 2.0). It simultaneously encodes three partially indistinguishable signals: attained biological age, length of follow-up in the registry (older participants have accumulated more diagnoses), and the era of parental exposure — individuals born earlier tend to have parents with potentially higher doses.
Dose variables — ranks 2–5
Parental doses (mother_aggregated_dose, mean_parent_dose, father_aggregated_dose, max_parent_dose) occupy ranks 2–5. Higher parental doses are associated with a moderately elevated predicted risk. The signal is robust but substantially weaker than the demographic effect.
5. Ablation Study
An ablation study is a methodology in which features are removed one at a time to quantify their contribution to prediction quality. Two trials were run: a model without birth_year — isolating the demographic confounder; and a model using only dose-related features — a lower-bound estimate of the signal attributable solely to parental exposure.
Chapter I
−0.060 … −0.070
ΔAUC = AUC_full − AUC_without_birth_year
Chapter C
−0.067 … −0.113
A more pronounced effect for the rarer outcome
Dose features only · AUC
0.619
–
0.761
Lower bound of the predictive signal attributable solely to parental exposure (father dose, mother dose, max dose, mean dose)
Conclusion. Dose-related features carry an independent predictive signal (AUC > 0.61 with demographic variables fully excluded), yet birth year contributes a substantial confounding effect. Biological age, length of follow-up, and the era of parental exposure are indistinguishable within the current registry structure. Causal interpretation of the dose signal requires additional steps: age stratification and multiple imputation (MICE) for records with missing doses.