Forecasting seasonal allergic rhinitis through integrated analysis of social media and online drug sales data

Background Allergic rhinitis (AR) is a highly prevalent, seasonally variable disease that poses a growing public health challenge. Conventional surveillance based on clinic visits and surveys is slow and may miss self-medicating patients. Integrating environmental monitoring with digital traces such as search queries and online drug purchases may provide more timely insight into allergen exposure, symptom awareness, and medication demand. Methods We analyzed daily data for urban Beijing from March 1, 2022 to October 15, 2024, including pollen concentration (grains per 1000 mm 3 ), the Baidu search index for "allergic rhinitis," and online purchase rates of 4 oral second-generation H1-antihistamines (loratadine, desloratadine, cetirizine, levocetirizine) from the Meituan platform, standardized per 100,000 population. Cross-correlation functions were computed on pre-whitened series to assess lead–lag relationships. A single-mediator regression framework applied to pre-whitened data and estimated with Newey–West heteroskedasticity- and autocorrelation-consistent standard errors quantified the mediating role of the Baidu Index between pollen and antihistamine purchases. ARIMAX (1,1,1) models with different exogenous inputs (pollen only, Baidu Index only, both combined) and a naïve random-walk benchmark were used to forecast 1-, 2-, 3-, and 7-day demand; performance was evaluated using RMSE, MAPE, and Pearson correlation (PCC). Results All 3 series exhibited pronounced and recurrent seasonal patterns, with Baidu search activity rising in close temporal alignment with pollen peaks and antihistamine purchases typically lagging by 1–2 days. Pre-whitened cross-correlations remained moderate and statistically significant around lag 0, indicating genuine contemporaneous associations after removing the shared seasonal component. Mediation analysis showed that pollen concentration had a significant total effect on antihistamine purchase rates (β = 0.34, 95% CI [0.13, 0.54]), of which approximately 47% was transmitted indirectly via the Baidu Index (indirect β = 0.16, 95% CI [0.08, 0.25]). The ARIMAX (1,1,1) model integrating both pollen and Baidu Index achieved the best forecasting performance (1-day RMSE=3.80, MAPE=7.68%, PCC=0.89) and consistently outperformed single-predictor models and the random-walk benchmark across all horizons. Conclusion Pollen exposure influences antihistamine demand both directly and indirectly through public information-seeking behavior captured by the Baidu Index, and integrating environmental and digital data enables timely surveillance of seasonal allergic-rhinitis–related medication use.