Livestock-related microbial air pollution: Implementing random forest modelling to predict residential endotoxin exposure

Abstract

Ambient endotoxin, a key component of livestock-related bioaerosols, is associated with health effects, highlighting the need for improved exposure assessment in rural populations. Previous studies relied on either resource-intensive deterministic dispersion models or more practical stochastic methods. So far, land use regression (LUR) is the only stochastic approach evaluated for ambient endotoxin. We assess random forest (RF) models as an alternative for predicting residential endotoxin exposure, hypothesising that its ability to capture complex, non-linear relationships will improve performance over LUR. RF models were trained using data from the Dutch VGO research programme (Livestock Farming and Neighbouring Residents' Health Study), with repeated measurements at 61 residences (236 total) and spatial livestock-related predictors. Model performance was assessed with 10-fold cross-validation (R2 and RMSE), and variable importance was assessed with mean decrease in impurity (MDI) and Shapley Additive Explanations (SHAP). RF slightly outperformed LUR at the simplest predictor level (R2: 0.19 vs 0.10), performed comparably at the intermediate level, and was outperformed by LUR at the most detailed level (R2: 0.24 vs 0.32). Both approaches tended to underestimate concentrations at high-exposure sites, indicating limitations in capturing the upper tail of the exposure distribution. Pig, poultry, and cattle metrics were consistently key predictors; SHAP revealed non-linear links with endotoxin, highlighting RF's ability to capture complex patterns that may be overlooked by linear approaches. Overall, RF provides a viable alternative to LUR in data-constrained settings where detailed predictor data are unavailable, but does not demonstrate a general performance advantage when detailed predictor data are available.