Beyond Residential Ambient Concentrations: Quantifying Exposure Error and Advancing Personal PM_2.5 Prediction with a...

Area of Science:

• Environmental Health Sciences
• Epidemiology
• Data Science

Background:

• Accurate personal fine particulate matter (PM2.5) exposure assessment is vital for epidemiological studies.
• Conventional ambient air quality data often lead to significant exposure misclassification.
• Existing methods struggle with scalability and precision in large-scale population studies.

Purpose of the Study:

• To quantify the errors associated with using ambient air quality data as proxies for personal PM2.5 exposure.
• To develop and validate a scalable modeling framework for predicting personal PM2.5 exposure using accessible data.
• To enhance the accuracy of exposure estimates in large epidemiological cohorts.

Main Methods:

• A panel study involving 12 adults across three Chinese cities, collecting 4571 person-hours of personal PM2.5 measurements.
• Comparison of personal measurements against three ambient data sources to quantify relative errors.
• Development of an integrated modeling framework using ambient concentrations, meteorological data, and personal characteristics, employing machine learning algorithms (Random Forest) with hyperparameter tuning and cross-validation.

Main Results:

• Substantial discrepancies were found between personal and ambient PM2.5 exposure, with daily average relative errors ranging from 39% to 48%.
• The developed Random Forest model, utilizing daily monitoring-station data, achieved high predictive performance (R² = 0.87).
• SHAP analysis confirmed ambient PM2.5 as the primary predictor, with personal traits and meteorological factors also showing significant contributions.

Conclusions:

• The study provides a validated, end-to-end modeling framework that significantly refines personal PM2.5 exposure estimation beyond traditional ambient proxies.
• This standardized workflow offers a scalable solution for improving the accuracy of exposure data in large-scale air pollution health research.
• The findings underscore the importance of moving beyond ambient data to reduce exposure misclassification and enhance the validity of epidemiological findings.