Performance of Prediction Algorithms for Modeling Outdoor Air Pollution Spatial Surfaces

Area of Science:

• Environmental Health
• Epidemiology
• Geospatial Analysis

Background:

• Land use regression (LUR) models are common for estimating air pollutant concentrations.
• Traditional linear regression methods face criticism for inflexibility and inability to handle predictor interactions or correlations.
• Ultrafine particles (UFPs) pose health risks, necessitating accurate exposure assessment models.

Purpose of the Study:

• To evaluate various modeling approaches for estimating long-term UFP concentrations.
• To compare the predictive performance of machine learning algorithms against traditional multivariable methods.
• To assess the importance of external validation data in model evaluation.

Main Methods:

• Utilized two training datasets: mobile UFP measurements (8200 segments) and short-term stationary measurements (368 sites).
• Evaluated model precision and bias using an independent external dataset (42 sites) for long-term average exposure estimation.
• Compared machine learning algorithms (bagging, random forest) with multivariable methods (stepwise regression, elastic net).

Main Results:

• Higher training R-squared did not guarantee higher test R-squared, highlighting the need for external validation.
• Machine learning models trained on mobile data explained 38-47% of external UFP concentrations.
• Multivariable methods (stepwise, elastic net) explained 56-62% of external UFP concentrations when trained on mobile data.
• Machine learning models trained on short-term stationary data showed modest improvements over linear and regularized regression.

Conclusions:

• The type of training data significantly influences the predictive ability of air pollution models.
• Traditional multivariable regression methods demonstrated superior performance over machine learning for UFP prediction using mobile measurement data.
• External validation is essential for reliably comparing the performance of different air pollution modeling techniques.