Evaluating the indoor radon concentrations in the Swedish building stock using statistical and machine learning

Abstract

Exposure to excessive indoor radon causes around 500 lung cancer deaths in Sweden annually. However, until 2020, indoor radon measurements were only conducted in around 16% of Swedish single-family houses and 17% of multifamily houses. It is estimated that approximately 16% of single-family houses exceed the indoor radon reference level of 200 Bq/m3, and the corresponding situation in multifamily houses is unknown. Measuring indoor radon on an urban scale is complicated and costly. Statistical and machine learning, exploiting historical data for pattern identification, provides alternative approaches for assessing indoor radon risk in existing dwellings. By training MARS (Multivariate Adaptive Regression Splines) and Random Forest (RF) regression models with the data labels from the radon measurement records in the Swedish Energy Performance Certification registers, property registers, soil maps, and the radiometric grids, the correlations between response and predictive variables can be untangled. The interplay of the key features, including uranium and thorium concentrations, ventilation systems, construction year, basements, and the number of floors, and their impact magnitudes on indoor radon concentrations, are investigated in the study. The regression models tailored for different building classes were developed and evaluated. Despite the data complexity, the RF models can explain 28% of the variance in multifamily houses, 24% in all buildings, and 21% in single-family houses. To improve model fitting, more intricate supervised learning algorithms should be explored in the future. The study outcomes can contribute to prioritizing remediation measures for building stocks suspected of high indoor radon risk.