Identifying Predictors of Spatiotemporal Variations in Residential Radon Concentrations across North Carolina Using Machine Learning Analytics

Zhenchun Yang; Lauren Prox; Clare Meernik; Yadurshini Raveendran; David Press; Phillip Gibson; Amie Koch; Olufemi Ajumobi; Jeffrey Clarke; Ruoxue Chen; Junfeng Jim Zhang; Tomi Akinyemiju

doi:10.1016/j.envpol.2024.125592

Identifying Predictors of Spatiotemporal Variations in Residential Radon Concentrations across North Carolina Using Machine Learning Analytics

Environ Pollut. 2025 Jan 4:125592. doi: 10.1016/j.envpol.2024.125592. Online ahead of print.

Authors

Affiliations

¹ Duke Global Health Institute, Durham, NC 27708, United States.
² Nicholas School of the Environment, Duke University, Durham, NC 27708, United States.
³ Department of Population Health Sciences, Duke University, Durham, NC 27708, United States.
⁴ Duke Cancer Institute, Duke University, Durham, NC 27708, United States.
⁵ North Carolina Department of Health and Human Services, Raleigh, NC 27612, United States.
⁶ Duke School of Nursing, Duke University, Box 3322, Durham, NC 27710.
⁷ Duke Global Health Institute, Durham, NC 27708, United States; Nicholas School of the Environment, Duke University, Durham, NC 27708, United States.
⁸ Department of Population Health Sciences, Duke University, Durham, NC 27708, United States; Duke Cancer Institute, Duke University, Durham, NC 27708, United States.

PMID: 39761718
DOI: 10.1016/j.envpol.2024.125592

Abstract

Radon is a naturally occurring radioactive gas derived from the decay of uranium in the Earth's crust. Radon exposure is the leading cause of lung cancer among non-smokers in the US. Radon infiltrates homes through soil and building foundations. This study advances methodologies for assessing residential radon exposure by leveraging a comprehensive dataset of 126,382 short-term (2-7 days) radon test results collected across North Carolina from 2010 to 2020. Employing a combination of linear regression and advanced machine learning techniques, including random forest models. Analysis through linear regression, linear mixed-effects models (LME), and generalized additive models (GAM) using the first-time tested radon levels reveals that elevation, proximity to geological faults, and soil moisture are pivotal in determining radon concentration. Specifically, elevation consistently shows a positive relationship with radon levels across models (linear regression: β=0.12, p<0.001; LME: β=0.17, p<0.001; GAM: β=0.11, p<0.001). Conversely, the distance to geological faults negatively correlates with radon concentration (linear regression: β=-0.11, p<0.001; LME: β=-0.06, p<0.001; GAM: β=-0.07, p<0.001), indicating lower radon levels further from faults. Using the random forest model, our study identifies the most influential environmental predictors of first-time tested radon levels. Elevation is the most influential variable, followed by median instantaneous surface pressure and soil moisture in the upper 10 cm layer, illustrating the significant role of geological and immediate surface conditions. Additional important factors include precipitation, mean temperature, and deeper soil moisture levels (40-200 cm), which underscores the influence of climate on radon variability. Root zone soil moisture and the Normalized Difference Vegetation Index (NDVI) also contribute to predicting radon levels, reflecting the importance of soil and vegetation dynamics in radon emanation. By integrating multiple statistical models, this research provides a nuanced understanding of the predictors of radon concentration, enhancing predictive accuracy and reliability.