Integrating GPS and Low-Cost Sensors to Enhance Personal PM_2.5 Exposure Modeling for Schoolchildren

Accurately estimating personal PM_2.5 exposure remains challenging due to limited data on individual spatiotemporal variability of PM_2.5 concentrations. This study developed a systematic framework to improve microenvironmental exposure (ME) models for estimating daily PM_2.5 exposure. A total of 918 person-days of PM_2.5 measurements were collected from schoolchildren in northern Taiwan over one year using portable PM_2.5 sensors equipped with GPS modules. Pre-deployment QA/QC calibration against a reference Met One BAM-1020 monitor showed high measurement accuracy (mean R² = 0.996; regression slope = 0.962). Time-location data obtained from sensors, GPS tracking, and questionnaires were used to develop six time-weighted average ME models. Personal PM_2.5 exposure exhibited substantial within-subject variability, as indicated by a low intraclass correlation coefficient (ICC = 0.125). Among the evaluated microenvironments, school had the highest mean PM_2.5 concentrations (22.7 μg/m³). The conventional ME model limited predictive performance and a narrower predicted exposure distribution. The fully GPS-resolved model (Model 6) performed best (R² = 0.619; accuracy = 74.0%); however, a simplified model using only GPS-identified home and school locations (Model 5) performed nearly equivalently (R² = 0.616; accuracy = 71.7%), indicating that the incremental gain from full GPS resolution was minimal. Most of the achievable improvement was therefore attributable to correctly characterizing the two dominant microenvironments rather than resolving every visited location. The proposed ME model reproduced the within- and between-subject variance structure observed in measured personal exposure. These results reflect internal validation against concurrently measured personal exposure; whether the models improve exposure ranking or reduce misclassification in exposure-health analyses was not evaluated and remains to be assessed in the LIGHTS cohort.