Radon, a naturally occurring gas originating from the ground, varies in concentration depending on geological and environmental factors. Radon-prone areas, as defined by the International Commission on Radiological Protection, exhibit significantly higher radon levels compared to other regions. Since measuring radon in every building is economically and logistically infeasible, predictive models offer a valuable alternative for assessing indoor radon levels. Using radon database provided by the Swiss Federal Office of Public Health (FOPH), extended with other data, this study evaluated four predictive methods: multiple linear regression, logistic regression, random forest regression, and random forest classification. These models incorporated diverse datasets, including geological, climatological, and building characteristics. Results revealed that random forest classification was the most effective, correctly predicting indoor radon levels above or below the 300 Bq/m3 reference threshold in 85% of cases. Random forest regression and logistic regression performed moderately, explaining 32% and 20% of variance, respectively, while multiple linear regression explained only 16% of the variance. Significant predictors included geology, building age, floor level, and foundation type, were consistent across methods, but also the previous literature. Predicting binary variables (above or below the reference level) proved more accurate than continuous radon level predictions. This study highlights the potential of machine learning methods, particularly random forest classification, to inform radon-prone area identification and guide public health interventions.
