Digital Prescription-Based Influenza Activity Forecasting in Jiangxi Province, China: Comparative Modeling Analysis...

Traditional laboratory-confirmed influenza surveillance involves a 1- to 2-week reporting delay and captures only patients who have sought care and received a diagnosis, limiting early warning. Digital prescription data have been shown to signal influenza activity early, but their predictive value for forecasting remains unclear.

We aimed to evaluate the predictive value of digital prescription data for influenza forecasting by assessing lead times, contributions across forecasting horizons, and added value when combined with other real-time sources.

Using daily data for Jiangxi Province, China (2022-2024), we characterized temporal relationships for 13 multisource indicators using prewhitened cross-correlation and performed Granger-based predictor screening restricted to the 2022 to 2023 development period. Incidence was then forecast 1 to 14 days ahead through expanding-window rolling-origin validation trained on 2022 to 2023 data and tested on 2024 data. Six models (autoregressive integrated moving average [ARIMA], seasonal ARIMA [SARIMA], Prophet, least absolute shrinkage and selection operator [LASSO], random forest, and explainable boosting machine) plus a persistence baseline were compared across 3 exogenous-input scenarios (no external input, digital prescriptions alone, and selected sources). Each indicator was also forecast from its own history and compared with observed incidence.

After prewhitening, the digital prescription rate led influenza incidence by 14 days, whereas the online search index lagged incidence by 1 day, indicating near-synchronous tracking. In the development period Granger analysis, digital prescription rate, online search index, and temperature showed significant predictor-to-incidence associations after false discovery rate (FDR) correction (FDR-adjusted P<.001, P=.003, and P=.008, respectively), whereas nitrogen dioxide showed a marginal association (FDR-adjusted P=.11) and was additionally included as an environmental covariate. At 1 day, performance was similar across models and input scenarios (R²=0.925-0.961; persistence R²=0.948), but differences emerged at longer horizons. SARIMA performed best with the digital prescription rate as the sole exogenous input, and LASSO performed best with 4-source input, both reaching a mean R² of 0.766 and root mean squared error (RMSE) of 0.240 across the 14 horizons. At 14 days, single-source SARIMA achieved an R² of 0.510 (RMSE 0.318), and multisource LASSO achieved an R² of 0.551 (RMSE 0.305), a gain of 0.064 in R² and a 6.2% reduction in RMSE over LASSO without exogenous input. Forecasts of the digital prescription rate showed the closest temporal consistency with observed influenza incidence (mean r=0.897 vs 0.866 for the online search index).

Digital prescriptions provide an early signal of provincial influenza activity, preceding routine reported incidence by approximately 2 weeks in this setting. Multisource input can further improve forecasting when paired with models capable of selecting informative signals, but indiscriminate integration does not necessarily help. This early signal may support earlier preparedness, including antiviral planning and health care resource allocation.