Predictive Modeling of Indoor Radiation Doses from Building Material Properties using Machine Learning: A Case Study from Delta State, Nigeria

Abstract

The use of naturally occurring radioactive materials (NORMs) in building construction can lead to potential long-term health risks for occupants from chronic low-dose radiation exposure. The study developed and validated five machine learning models to predict indoor radiation doses using building material properties based on 72 samples collected from Northern Delta State, Nigeria. Activity concentrations of potassium-40 (⁴⁰K), uranium-238 (²³⁸U) and thorium-232 (²³²Th) were measured using gamma-ray spectrometry. Models for predicting annual effective dose (AED) were trained using Multiple Linear Regression (MLR), Ridge Regression (RR), Random Forest (RF), Gradient Boosting (GB), and Support Vector Regression (SVR). MLR (R²= 1.000, RMSE = 1.27 × 10^-8 mSv/y) and RR (R² = 1.000, RMSE = 7.44 × 10^-6 mSv/y) showed near-perfect prediction accuracy, while RF (R² = 0.915) and GB (R² = 0.929) also exhibited excellent performance. The feature importance analysis revealed ²³⁸U as the dominant predictor (71.00%) followed by ²³²Th (19.95%) and ⁴⁰K (5.12%). The category- and origin-specific models achieved perfect accuracy (R² = 1.000) and offer practical tools for rapid radiological screening. The results show that linear models can predict indoor radiation doses with almost zero error due to the inherent linearity of the physics of dose calculation. However, the interpretation of these findings is limited by study limitations, such as the relatively small sample size (N=72) and the deterministic nature of the dose formula. This can be used in construction industries for efficient regulatory compliance and material selection.