Time-dependent radiological risk assessment following a severe NPP accident using Monte Carlo simulation, BEIR VII,...

Nuclear reactor accidents can result in prolonged radiation exposure with complex and uncertain health consequences. Existing studies often focus on dose estimation without fully linking time-dependent exposure dynamics to population-specific cancer risks. This study presents a comprehensive probabilistic framework for long-term radiological risk assessment, integrating time-resolved dose modeling, stochastic Monte Carlo simulation, BEIR VII-based epidemiological risk functions, and interpretable artificial intelligence. Radiation dose evolution was derived from HotSpot simulations under a severe VVER-1000 accident scenario at 200 km, capturing cumulative Total Effective Dose Equivalent (TEDE) over a 10-year period. Approximately 8 million virtual individuals were simulated to account for variability in environmental conditions, structural shielding, and human behavior. A deep learning surrogate model, optimized via Hyperband, achieved high predictive accuracy (R² ≈ 0.994) and a 60% reduction in prediction error, with a 15× computational speedup over conventional methods. Results reveal a pronounced age-dependent risk structure: children (0-10 years) show the highest Excess Relative Risk (ERR), particularly for leukemia (mean > 55; upper bounds >300), and females consistently exhibit higher and more variable risk across most cancer types. Risk distributions narrow with age, reflecting reduced sensitivity and stochastic variability in adults. SHAP-based sensitivity analysis confirms age as the dominant risk driver, with behavioral factors critical for leukemia and structural shielding more influential for solid cancers. The proposed framework provides a physically consistent, computationally efficient, and fully probabilistic approach, bridging physics-based modeling with data-driven inference. It enables real-time decision support and informs long-term radiological protection strategies.