Global distribution and evolutionary trends of the PM2.5 health burden predicted with a Geographically Neural Network Weighted Regression model.

PM2.5 pollution, driven by various human activities, has become a significant global threat to both environmental and public health, yet accurately predicting its health burden remains challenging due to the complex interplay of multiple influencing factors. To overcome the limitations of traditional models, including their inability to capture spatial non-stationary relationships, weak spatial interpretability, and dependence on densely distributed training data, this study proposes a hybrid forecasting framework based on Geographical Neural Network Weighted Regression. Based on global PM2.5 exposure and health data from 2000 to 2021, we integrated graph neural networks with spatially adaptive regression techniques to construct a hybrid model. This model effectively captures the spatial heterogeneity of PM2.5-related health burden and reveals its driving factors. The results showed that PM2.5-related mortalities were estimated to have increased by 57.25 % by 2021 and are projected to rise by an additional 12.84 % by 2030, with hotspots concentrated in low- and middle-income countries. The Health index was identified as the most influential factor (34.87 %), although interactions among Health, Income, and Material indicators varied by country. This study provided an interpretable model to support early warning systems and evidence-based public health strategies.