Estimating national indoor radon exposure at a high spatial resolution – improvements by a machine learning based probabilistic approach

Accurate knowledge of indoor radon exposure is vital information for assessing radon-induced health effects, identifying radon prone areas or estimating the number of people affected by the exceedance of a specific radon concentration in a given area.

Large-scale measurement campaigns are usually the tool of choice for determining regional or national indoor radon exposure. These campaigns need to be representative in terms of all relevant factors governing indoor radon exposure (e.g., geogenic radon availability, distribution of people across floor levels, building types) for providing an unbiased estimate. In practice, creating a fully representative sample of the population is hardly achievable due to the multitude of relevant factors which cannot be fully controlled by sampling design. Further, estimating indoor radon exposure at a high spatial resolution (district or municipality scale) requires a large number of measurements which increases the financial and logistic effort dramatically. Therefore, predictive models are widely applied as a complementary tool for exposure assessment by utilizing available information on the relevant variables that determine indoor radon. However, these models are usually only able to explain a certain amount of the observed variability due to the absence of some relevant information (building-specific data on air tightness, ventilation rates etc.). As a consequence, model-based assessments tend to underestimate the true variability of indoor radon. 

In this study, we present a probabilistic approach that overcomes this shortcoming and intends to give a more realistic estimate of the true indoor radon distribution at several spatial resolutions. Our approach consists of the following steps:

1) fitting a random forest model utilizing 12 predictors to ~14,000 full-year indoor radon measurements in residential buildings in Germany;

2) predicting a range of quantiles of the expected indoor radon distribution for each floor level of each German residential building;

3) fitting a lognormal distribution to the estimated quantile data to approximate the building and floor level specific probability density function (PDF);

4) random sampling from this PDF with a sample size proportional to the population distribution;

5) aggregating results on several spatial scales. 

The benefits of this approach are 1) to allow an accurate exposure assessment even if surveys were not fully representative concerning the main controlling factors by utilizing high-resolution information on the spatial distribution of these factors via predictive models. 2) with a given amount of measurements, exposure distribution can be estimated at a much higher spatial resolution compared to basic aggregate statistics.