Large-scale spatiotemporal deep learning predicting urban residential indoor PM_2.5 concentration

Area of Science:

• Environmental Health
• Data Science
• Public Health

Background:

• Indoor PM2.5 pollution is a significant global cause of mortality and morbidity.
• Large-scale monitoring of indoor PM2.5 is difficult, hindering population-level risk assessment.
• Existing models for indoor PM2.5 lack generalizability and often provide only single-point predictions.

Purpose of the Study:

• To develop a generalized, easy-to-use model for predicting indoor PM2.5 concentrations and their spatiotemporal variations globally.
• To assess population-level exposure and associated health risks of indoor PM2.5 pollution.
• To overcome limitations of existing machine learning models in predicting indoor PM2.5.

Main Methods:

• Developed a Bayesian neural network (BNN) model to predict daily average urban residential PM2.5 concentrations.
• Utilized comprehensive nationwide sensor-monitoring data from China.
• Validated the model using 10-fold cross-validation, achieving an R² of 0.70.

Main Results:

• The BNN model demonstrated strong predictive performance with a mean absolute error of 9.45 μg/m³ and root-mean-square error of 13.3 μg/m³.
• Achieved 95% prediction interval coverage of 85%, indicating reliable uncertainty quantification.
• Modeled population-weighted annual indoor PM2.5 concentration in China for 2019 was 22.8 μg/m³, exceeding WHO standards.

Conclusions:

• The developed BNN model offers a robust method for large-scale prediction of indoor PM2.5 concentrations.
• The model's population-level validity is valuable for managing indoor air pollution and related health risks.
• Findings highlight the urgent need for interventions against high indoor PM2.5 levels, especially in China.