Correlation Analysis of Air Pollutants and Human Diseases Using Explainable AI for Health Outcomes

Abstract

Air pollution is a prominent global health risk factor responsible for millions of premature deaths annually, with its impact spanning cardiovascular, respiratory, and malignant disease categories. Predicting the health impact of air pollution across diverse national contexts requires modelling complex non-linear interactions between multiple pollutant species and disease-specific outcomes, a task that conventional single-pollutant epidemiological frameworks cannot adequately address. To address this, the present research proposes an Explainable AI-enabled predictive framework integrating EDGAR v8.1 global emission inventories with WHO Global Health Observatory health metrics across 177 countries, 9 pollutant species, and 6 disease categories for the period 2010-2019. Two novel deep learning architectures are introduced: RBN-BiLSTM, which extends the standard BiLSTM with Batch Normalisation and L2 regularisation for short emission panel sequences, and T-RBN-BiLSTM, which further incorporates a temporal attention mechanism to dynamically weight emission years by their predictive relevance. The models are evaluated under both random split and strict temporal split protocols and compared against classical ML baselines, including XGBoost and Extra Trees. The proposed T-RBN-BiLSTM achieves 95.8% temporal accuracy and an AUC of 98.2%, outperforming baseline BiLSTM by 12.6 percentage points. SHAP-based per-disease attribution is applied across all six WHO disease categories to identify pollutant-specific health impact drivers. Results demonstrate that agricultural NH3 and biomass-burning organic carbon are the dominant positive health impact drivers across cardiovascular and respiratory diseases, while NOx and SO2 act primarily as economic confounding signals. PM2.5 ranks last due to multicollinearity with co-emitted combustion species, highlighting the importance of multi-pollutant frameworks for accurate policy attribution. Future work includes incorporating sub-national resolution and socioeconomic covariates to further improve the predictive and explanatory scope of the framework.