Interpretable Machine Learning for Air Pollution and Respiratory Health Prediction: A Socioeconomic Subgroup Analysis

This study evaluates an interpretable machine learning framework to predict respiratory disease rates and air-quality status using a dataset of 14,100 weekly country-level records. The researchers compared nine regression and nine classification models using nested cross-validation to ensure rigorous hyperparameter tuning and performance estimation. To address the risk of target leakage, the study performed a sensitivity analysis on air-quality classification by comparing models with and without PM2.5 data. Finally, the authors used SHAP (SHapley Additive exPlanations) values to interpret model decisions and conducted subgroup analyses across different income levels and geographic regions to identify variations in predictive patterns.

The analysis confirms that PM2.5 concentration is the dominant signal for predicting respiratory disease rates, with linear and regularized linear models performing best. In the air-quality classification task, models achieved high accuracy when PM2.5 was included, but performance dropped significantly when it was removed, demonstrating a strong dependence on pollutant-related information. SHAP analysis revealed that when PM2.5 is absent, the models shift their reliance toward socioeconomic and meteorological variables, such as GDP per capita, precipitation, and healthcare access. Subgroup analysis indicated that while aggregate prediction errors were similar across income levels, the specific contribution of PM2.5 to model predictions was notably stronger in lower-middle-income countries.

This research highlights that high predictive accuracy in climate-health models can sometimes mask a reliance on a single dominant proxy variable rather than a complex understanding of environmental health drivers. By demonstrating the use of SHAP values and sensitivity analysis, the paper provides a template for researchers to audit their models for target-proxy dependence and to uncover potential disparities in how environmental factors influence health predictions across different socioeconomic contexts. This is essential for building robust, transparent, and equitable public health surveillance tools.