2611

Background. Traditional air pollution epidemiologic models use ground-based monitors and/or satellite estimates of ozone and PM_2.5 concentrations as surrogates of exposure. It is possible that crowd-sourced or online data may be more effective at reflecting population exposures than these traditional indicators. Here, we constructed models to “nowcast” observed elevated pollution levels, using online search queries to create a Crowd-Generated Air Pollution (CGAP) metric and compared the output to monitoring data. We then used the CGAP metric to examine short-term associations of ED visits for pediatric asthma. Methods. A search term Dictionary Learner-Long-Short Term Memory (DL-LSTM) composite model was developed to combine meteorological data, air pollution measures, and pollution-related search data for Atlanta, and daily CGAP index was generated for ozone and PM_2.5 from 2007 to 2008. Daily counts of ED visits for asthma among children aged 5 to 17 were collected as well. Using quasi-Poisson generalized linear models, we assessed short-term associations of measured ambient concentration and modeled CGAP index with ED visits. Results. Both ambient ozone and PM_2.5 levels were associated with asthma ED visits. We observed similar patterns for indices from DL-LSTM model trained by meteorological data, air pollution measures, and search data (r = 0.88, p < 0.0001 for ozone; r = 0.71, p < 0.0001). The quasi-Akaike information criterion (AIC) of the modeled CGAP index was similar to or lower than that of ambient levels (3179.399 vs. 3173.754 for ozone; 3177.887 vs. 3198.606 for PM_2.5). Conclusion. Our findings provided preliminary validation of the CGAP metric, based on a DL-LSTM algorithm. Results using CGAP were comparable to findings using stationary monitoring data for air pollution. The modeled CGAP index showed promise as a means of potentially providing a sensitive, comprehensive predictor of short-term changes in pediatric asthma ED visits due to urban air pollution exposures.