Construction of air quality health index based on multi-pollutant Bayesian kernel machine regression and deep...

Efficient characterization and early warning of health risks associated with air pollution are critical issues in public health management. However, the traditional AQHI (Air Quality Health Index) suffers from inconsistent criteria for pollutant selection, insufficient consideration of environmental covariates, and an inability to reflect interactions between pollutants.

Based on multi-source data from Wuhan (2017-2019), we constructed an optimized AQHI using machine learning and BKMR (Bayesian Kernel Machine Regression) methods. The performance of the new index was compared with that of traditional methods in characterizing health risks across different populations. Additionally, a deep learning-based early warning model was developed to improve the identification capability for high-risk pollution events.

Temperature, NDVI, and other variables were selected as important covariates, and an interaction between PM_2.5 and NO₂ was identified. Compared to Standard-AQHI and CRI-AQHI, the new index demonstrated superior performance in characterizing non-accidental hospitalization risks (3.80%, 95% CI: 3.11%-4.49%), particularly among cardiovascular disease patients (4.74%, 95% CI: 4.10%-5.40%) and the elderly population (4.82%, 95% CI: 4.09%-5.56%). The optimized early warning model showed significant improvements (R² increased from 0.76 to 0.81; F1-score increased from 0.32 to 0.79).

The enhanced AQHI demonstrated superior performance in pollutant interaction analysis, risk characterization, and high-risk identification. These findings provide a scientific basis for formulating environmental health policies, optimizing public health warning systems, and enhancing risk communication.