Near real-time modeling of daily asthma emergency department visits syndrome as a function of air pollution, weather, and other environmental risk factors in New York City, 2010-2019

Abstract

Background: Research has shown that asthma morbidity is associated with multiple temporally-varying environmental risk factors including air pollution, weather, pollen, and viral infections. Predicting its temporal variation near real-time is useful for situational awareness as the relative importance of these risk factors changes across seasons.Methods: We developed a regression model to predict daily asthma emergency department (ED) visits based on chief complaint text and diagnosis code (i.e., syndrome) available from previous day since 2010 at the New York City Health Department. The 2010-2018 data were used to develop a quasi-Poisson model for a 121-day moving time window for each day of 2019 for ages 5-17, with predictors including: day-of-year; day-of-week; temperature; citywide average fine particles (PM2.5) and ozone (warm season only); ED visits syndromes for allergy (proxy for spring pollen impact) and flu; school-opening date (proxy for fall rhinovirus infections); and holidays, considering non-linear distributed lags when multi-day impacts were observed. Point estimates and prediction intervals are produced along with the observed values in a dashboard to facilitate daily assessment.Results: The model goodness of fit varied across seasons, with the best fit occurring in the fall season (e.g., % deviance explained up to 85%) and the worst in the spring (e.g., ~ 40%). Daily absolute percent errors, with a median of 15% for the 2019 prediction period, were also higher in the spring, because the peak tree pollen impacts were not fully captured. PM2.5 associations were mostly limited to cold months, with diminished predictive power during the warm period when ozone was a better predictor.Conclusion: The dynamic near real-time prediction model for asthma ED syndrome for young age group developed in this study provides a useful tool for evaluating the impacts of current events. Further model improvement will consider spatially-resolved predictors (e.g., satellite-derived air pollution data).