B30-03 Comparison of Machine Learning Derived and Conventional Census Tract Level Pm2.5 Exposure Datasets in Relation to Veteran Emergency Visits in California During 2016-2018

Abstract Rationale Fine particulate matter (PM2.5) is a major driver of respiratory disease, and daily fluctuations have been linked to respiratory emergency (ER) visits. PM2.5 datasets differ in construction and data sources: EPA-modeled averages are convenient but rely on sparse ground monitors, whereas machine-learning (ML) products fuse satellite, monitor, and model inputs to better capture spatial and temporal variability at the same ZIP Code scale. Evaluating whether these methodological differences influence observed exposure-response relationships, this preliminary study compares EPA and ML PM2.5 datasets in relation to respiratory emergency visits. Methods Using a subset of the Airborne Hazards and Open Burn Pit Registry, we applied a time-stratified case-crossover design to assess associations between PM2.5 and respiratory ER visits in California (2016-2018). Each case day was matched to control days within the same ZIP Code Tabulation Area, day of week, month, and year. Daily PM2.5 exposures from EPA and ML datasets were used to estimate lagged (0-5 days) and cumulative (0-2, 0-5 days) effects. Conditional logistic regression was used to estimate odds ratios per +10 µg/m³, with secondary analyses restricted for high PM2.5 events (≥25 µg/m³, 0-2 days). Results Among 370 respiratory ER visits across 232 ZCTAs in California (2016-2018), time-stratified case-crossover analyses showed positive short-term associations between PM2.5 and ER visits. For the continuous exposure models, effects peaked at lag 1 day for both datasets: ML-derived PM2.5 yielded an odds ratio (OR) = 1.22 (95% CI: 0.99-1.48) per +10 µg/m³, while EPA-derived PM2.5 yielded OR = 1.12 (95% CI: 0.96-1.29). Cumulative exposure windows indicated consistent, though modest, associations (ML 0-2 days: 1.20 0.97-1.48; EPA 0-2 days: 1.11 0.95-1.31). When restricting to high-pollution events (≥ 25 µg/m³, 0-2 day mean), point estimates increased, ML OR is 1.67 (0.71-3.93) and EPA OR is 1.39 (0.69-2.79), suggesting stronger but less precise effects under extreme smoke or pollution episodes. Conclusion Short-term PM2.5 increases were associated with higher odds of respiratory ER visits, with peak effects within one day of exposure. The ML-derived PM2.5 detected stronger and more temporally responsive associations than the EPA dataset, indicating improved sensitivity to localized pollution variability and transient wildfire smoke plumes. Although confidence intervals included the null, the consistent direction and magnitude of effects across models suggest meaningful patterns that warrant confirmation in larger samples. Ongoing analyses using a statewide Veteran cohort will enhance power, refine subgroup analyses, and assess whether ML-derived enhancements further improve detection of population and season-specific vulnerabilities. This abstract is funded by: VA’s Military Exposure Research Program