Integrating spatial analytics and explainable machine learning to analyse multi-pollutant mortality burdens in Africa

Abstract Air pollution in Africa is a critical public health problem that has claimed many lives, especially in places where urbanisation and vehicle emissions are increasing. The present study used spatial analysis and a hybrid Genetic Algorithm–Machine Learning (GA–ML) approach to predict mortality rates, in a quest to minimise risks associated with this problem. The study employed varied datasets, in which SVR, Support Vector Regression-Generic Algorithm (SVR-GA), LightGBM, and LightGBM-GA were examined, using Egypt, Nigeria, Kenya and South Africa as a case study. In the spatial analysis, Nigeria and Egypt were identified as hotspots of mortality, whereas South Africa had moderate mortality rates. The SVR-GA algorithm registered superior results, producing an R 2 of 86.4%, MAE (0.034) and MSE (0.01), while the SVR had the second-best. SHAP analysis highlighted PM2.5 and CO as the dominant factors in building the models, highlighting the need for close monitoring of these pollutants. The model’s generalisation ability was improved by conducting the LOCO-CV and the country-wise validation processes. According to the results obtained, there is some level of positive spatial autocorrelation for deaths from air pollution, and hotspots and coldspots can be observed for the studied countries. Of all the tested models, SVR-GA performed the best, outperforming SVR, LightGBM, and LightGBM-GA. In terms of geographic validation, SVR-GA achieved the best results in predictive ability and stability across the countries studied, confirming its high generalisation ability. Using the Shapley Additive exPlanation (SHAP) technique, it was established that PM₂.₅ and CO were the most crucial predictors of the models’ outputs. Moreover, the results were validated by the Taylor diagram and the Empirical Cumulative Distribution. In this way, the study contributes novel insights to environmental modelling through the integration of spatial statistics, advanced machine learning, metaheuristic optimisation, and a thorough evaluation. Furthermore, the hybrid GA-machine learning framework is a powerful and adaptable tool that can assist with data-driven environmental management and the realization of the Sustainable Development Goals relating to health, climate action, and sustainable urban development.