Seasonal allergic rhinitis (AR), also known as hay fever, is a common respiratory condition brought on by a range of environmental triggers. Previous work has characterised the relationships between community-level AR symptoms collected using mobile apps in two Australian cities, Canberra and Melbourne, and various environmental covariates including pollen. Here, we build on these relationships by assessing the skill of models that provide a next-day forecast of an individual's risk of developing AR and that nowcast ambient grass pollen concentrations using crowd-sourced AR symptoms as a predictor. Categorical grass pollen forecasts (low/moderate/high) were made based on binning mean daily symptom scores by corresponding categories. Models for an individual's risk were constructed by forward variable selection, considering environmental, demographic, behaviour and health-related inputs, with non-linear responses permitted. Proportional-odds logistic regression was then applied with the variables selected, modelling the symptom scores on their original five-point scale. AR symptom-based estimates of today's average grass pollen concentration were more accurate than those provided by two benchmark forecasting methods using various metrics for assessing accuracy. Predictions of an individual's next-day AR symptoms rated on a five-point scale were correct in 36% of cases and within one point on this scale in 82% of cases. Both outcomes were significantly better than chance. This large-scale AR symptoms measurement program shows that crowd-sourced symptom scores can be used to predict the daily average grass pollen concentration, as well as provide a personalised AR forecast.
Keywords: Allergic rhinitis; Citizen science; Modelling; Pollen; Symptom score.
Copyright © 2020 Elsevier B.V. All rights reserved.