Residential radon – Comparative analysis of exposure models in Switzerland

Residential radon exposure is a major public health issue in Switzerland due to the known association between inhaled radon progeny and lung cancer. To confirm recent findings of an association with skin cancer mortality, an updated national radon model is needed. The aim of this study was to derive the best possible residential radon prediction model for subsequent epidemiological analyses. Two different radon prediction models were developed (linear regression model vs. random forest) using ca. 80,000 measurements in the Swiss Radon Database (1994–2017). A range of geographic predictors and building specific predictors were considered in the 3-D models (x,y, floor of dwelling). A five-fold modelling strategy was used to evaluate the robustness of each approach, with models developed (80% measurement locations) and validated (20%) using standard diagnostics. Random forest consistently outperformed the linear regression model, with higher Spearman's rank correlation (51% vs. 36%), validation coefficient of determination (R² 31% vs. 15%), lower root mean square error (RMSE) and lower fractional bias. Applied to the population of 5.4 million adults in 2000, the random forest resulted in an arithmetic mean (standard deviation) of 75.5 (31.7) Bq/m³, and indicated a respective 16.1% and 0.1% adults with predicted radon concentrations exceeding the World Health Organization (100 Bq/m³) and Swiss (300 Bq/m³) reference values. Residential radon prediction models were developed and validated using a large number of measurements. Machine learning using random forest was found to perform substantially better than the more classical linear regression.