Exploring statistical and machine learning techniques to identify factors influencing indoor radon concentration

Area of Science:

• Environmental Science
• Public Health
• Geosciences

Background:

• Radon is a radioactive gas with carcinogenic effects, posing significant health risks primarily through indoor exposure.
• Indoor radon concentration (IRC) is influenced by various environmental and structural factors.
• Understanding these factors is crucial for mitigating public health risks associated with radon exposure.

Purpose of the Study:

• To investigate the combined effects of geology, pedology, and house characteristics on indoor radon concentration (IRC) in Romania.
• To identify the main predictors influencing indoor radon levels.
• To compare the performance of different statistical and machine learning models in predicting IRC.

Main Methods:

• Conducted 3132 passive radon measurements across Romania.
• Employed univariate statistics, artificial neural networks, and random forest regressor (RFR) for analysis.
• Utilized logistic regression and random forest classifier for binary classification of IRC.

Main Results:

• The Random Forest Regressor (RFR) model showed the best performance for continuous radon concentration (R²=0.14, RMSE=0.83).
• For discretized IRC (median 115 Bq/m³), logistic regression achieved an AUC-ROC of 0.61, and random forest classifier achieved 0.62.
• Key predictors identified include the presence of a cellar, construction period, altitude, and floor type.

Conclusions:

• Geology, pedology, and house characteristics are significant determinants of indoor radon concentration.
• Machine learning models, particularly RFR, offer robust methods for predicting radon levels.
• Identifying main predictors like cellar presence and construction period can inform targeted radon mitigation strategies.