Predictive and Explainable Machine Learning Models for Endocrine, Nutritional, and Metabolic Mortality in Italy Using Geolocalized Pollution Data

This study investigated the predictive performance of three regression models—Gradient Boosting (GB), Random Forest (RF), and XGBoost—in forecasting mortality due to endocrine, nutritional, and metabolic diseases across Italian provinces. Utilizing a dataset encompassing air pollution metrics and socio-economic indices, the models were trained and tested to evaluate their accuracy and robustness. Performance was assessed using metrics such as coefficient of determination (r2), mean absolute error (MAE), and root mean squared error (RMSE), revealing that GB outperformed both RF and XGB, offering superior predictive accuracy and model stability (r2 = 0.55, MAE = 0.17, and RMSE = 0.05). To further interpret the results, SHAP (SHapley Additive exPlanations) analysis was applied to the best-performing model to identify the most influential features driving mortality predictions. The analysis highlighted the critical roles of specific pollutants, including benzene and socio-economic factors such as life quality and instruction, in influencing mortality rates. These findings underscore the interplay between environmental and socio-economic determinants in health outcomes and provide actionable insights for policymakers aiming to reduce health disparities and mitigate risk factors. By combining advanced machine learning techniques with explainability tools, this research demonstrates the potential for data-driven approaches to inform public health strategies and promote targeted interventions in the context of complex environmental and social determinants of health.