Empirical research has found that distances ranging from 25 to 41 km have a good correlation of pollen measures [9, 24, 25]. In the United States, consistent pollen collection data stations are much sparser, with approximately 85 stations nationwide [26]. Using the information from the single pollen site available in Minnesota, our findings indicate that for evaluating thunderstorm asthma events, there is some predictive value of pollen levels at distances greater than previously tested. However, the measured effect diminishes with increasing distance. This reduced effect could be caused by misclassification or some other unmeasured factor. We found that the combined area of deciduous trees plus grassland is positively associated with the effect size of the relative risk of severe asthma on the day of thunderstorm events in 19 study locations, suggesting effect modification.

Our findings suggest that locations with plants similar to a remote pollen measurement location may have comparable daily pollen loads and provide predictive value for thunderstorm asthma research. In contrast, the linear decreasing effect size associated with greater distance from the central pollen site reflects an expected exposure measurement error that would plausibly lead to a biasing towards the null of health estimates at increasing distances [9, 24, 25].

Our a priori selection of deciduous trees and grassland land cover types is plausible based on previous literature demonstrating that plants, including ryegrass, or some tree species such as birch, olive, and grassland weeds, including pellitory [23], may have some involvement. However, this vegetation choice also lacks precision given the absence of additional species information and, more importantly, the nascence of the literature on pollen types associated with allergic-type thunderstorm asthma. Advancing the epidemiology of thunderstorm asthma requires a deeper understanding of what vegetative species most contribute to sub-pollen particle loads and how this can differ by time of year and geographic location. Limitations:

This study has several limitations. Exposure is ecological, and we cannot know the true exposure of any individual. Additionally, we cannot know if travel times to emergency departments affect utilization for less severe cases; however, all sites have a centrally located emergency department, and the population-weighted mean distance to an emergency room for each resident is 4.6 miles (2.0, 7.7) [27]. While we have explored differences in risk by age and sex during thunderstorm asthma, the data provided by the Minnesota Department of Health contains no information about race, ethnicity, or other social or demographic factors, and we could not examine their potential association with exposure or disease. Landcover data is a crude approximation of true pollen types. With only one single-point measurement for pollen, we cannot directly examine the predictive capacity of our measurement. We used a 15-mile radius area for each site for landcover measurement, but because of state boundaries, six sites are clipped and have smaller areas. This makes comparisons based on area percentages difficult as denominators differ and potentially undercount the total amount of square miles of land cover exposure. However, this approach best aligns with our health data as that is also not available for zip codes in neighboring states. We use a single year of landcover data to estimate landcover type, which could introduce additional bias or measurement error. Severe asthma is relatively rare, as are thunderstorm asthma events, which could introduce errors due to chance or lead to imprecise measurement. While our individual site models include ozone and PM2.5, monitoring sites, these are not equally distributed throughout the state, which could introduce confounding; however, prior research found no evidence that these variables were associated with the outcome [16]. The phenomenon of thunderstorm asthma is poorly understood, and there is much to be explored about the factors that contribute to thunderstorm asthma events, including exploring effects by individual plant species, and further exploration of how meteorologic factors such as rain, wind speed and storm type, may affect the process of sub-pollen particle creation and dispersal and modify the association between storms, pollen and asthma [28].

Conclusions Improved estimation of pollen for health research requires accurate modeling and measurement for exposure purposes. Current commercial pollen models are based on historical data and current weather [29, 30] and some incorporation of species-specific landcover data can improve the underlying data for prediction [31]. However, at this time, many pollen prediction models report poor correlation with ground-measured local pollen counts [32], and while computationally intensive models are improving, their results remain similar to commercial models [33].

Future researcher might combine vegetation landcover data [34] with NDVI (normalized difference vegetation index) to test whether pollen seasonality is indeed evidenced. Another approach that would provide greater precision in pollen estimation would be to add additional pollen collection stations. Current developments in automated pollen collection systems may support this in the near future [35], however, at this time, automated pollen measurement may allow faster counts but does not provide counts as accurate as more labor-intensive manual methods [36]. Our research suggests additional pollen measurements might allow for more accurate thunderstorm asthma prediction, and previous work suggests that a greater density of pollen measurements is required [9, 24, 25]. It is plausible that improvements in spatiotemporal modeling of daily pollen loads, similar to advances in air pollution exposure assessment [37, 38], will be a future data source for public health research. However, while our results suggest that land cover can improve prediction, improved modeling and forecasting is currently limited by the limited number of pollen stations with standardized data, imprecise data on land cover and plant speciation, as well as a need for more research investment to develop better informed models that account for regional variation in weather or climatic regimes [39].

Current automated systems that measure pollen show 84% concordance with ‘gold-standard’ human analysis [25, 40], allow faster reporting times with lower labor costs. Allergic rhinitis is costly, associated with pollen costs $4.5 billion per year in the United States [41] and in a cost-benefit analysis of a dense automated system in Bavaria (Germany), Buters et al. argue that a 0.1% reduction in health costs for allergies alone could offset the cost of a denser automated network in that setting [25, 42]. Several U.S. states are exploring more robust pollen measurement networks [43], but solutions for funding challenges and integration of regional data need to be addressed. Successful implementation of pollen modeling for prediction will require concurrent developments in local pollen measurement, a better understanding of the role of the landcover factors, and improvements in atmospheric modeling.

While there are no documented examples of epidemic thunderstorms in the U.S., there is evidence that climate change is driving a lengthening pollen season with increasing pollen loads [8], and studies suggest that higher temperatures may lead to more frequent thunderstorms in the United States [44]. While the impact of U.S. thunderstorm asthma events is modest regarding emergency room visits and healthcare burden, better data collection would allow an improved exploration of this work and provide a better estimation of tools to predict these events that may worsen in the future. This study highlights the importance of additional pollen measurement in the United States for further research into this specific health outcome and other important research areas, including studies of allergenic response and the spatiotemporal spread of pollen.