TransHealth: An Explainable AI Framework Combining Transformer Architectures and SHAP for Air Quality-Driven Health Risk Assessment

Air pollution poses critical global health challenges, necessitating accurate and interpretable prediction systems. We propose a hybrid framework combining Transformer encoders for spatiotemporal feature extraction, XGBoost for multi-class health impact classification, and SHAP for model explainability. Evaluated on real-world air quality data with five health impact categories (0-4), our model achieves 95.8% accuracy and microaverage AUC of 0.927. SMOTE successfully addresses class imbalance, yielding F1-scores of 0. 7 4 5 - 0. 7 9 3 for minority severe impact classes. SHAP analysis identifies PM2.5, AQI, and PM10 as primary drivers with critical thresholds, while revealing protective effects of meteorological factors. This interpretable framework bridges predictive accuracy and clinical utility for public health decision-making.