Bridging the Gap Between Data Reproduction and Prediction: The Impact of Feature Selection and Cross-Validation...

Reliable exposure assessment is vital for epidemiological research, but weaknesses in land-use regression (LUR) models undermine its validity. Using mobile ultrafine particle (UFP) data in Toronto, we compared LUR models trained under random, spatial, temporal, and spatiotemporal cross-validation (CV), with and without forward feature selection (FFS). Model hyperparameters and feature subsets were optimized within each CV scheme. Spatial CV folds were designed at fine scales to reflect UFP autocorrelation. Each approach was evaluated on a hold-out test set, across CV schemes, and against independent stationary backyard measurements. Models based on spatiotemporal CV coupled with FFS were able to reduce overfitting, improve generalization, and produce stable exposure surfaces. These surfaces avoided the spatial artifacts and exaggerated variable effects typically seen in models trained with random CV. Models tuned with random CV overfit, performed poorly on independent samples, and were sensitive to outliers. The average percentage error (APE) decreased from ∼217% for a model with random-CV to ∼79% with spatiotemporal CV and FFS. Our findings demonstrate that proper alignment of model design with the data's spatiotemporal structure and modeling objective ensures reliability, minimizes data reproduction, and enables true prediction.