Nonlinear Granger causality and LSTM-Transformer hybrid forecasting reveal meteorological drivers of influenza incidence in Gansu Province

Le résumé fourni par la source

Influenza poses a significant public health burden with distinct seasonality. In regions with pronounced climatic features like Gansu Province, China, traditional statistical models often fail to capture the nonlinear and lagged effects of meteorological drivers on influenza transmission. A deeper understanding of these complex relationships is essential for improving predictive accuracy. A Random Forest-based nonlinear Granger causality test was employed to systematically identify significant lagged predictive relationships between eight meteorological variables and monthly influenza incidence in Gansu Province (2004–2020). This approach accommodates complex, nonlinear dependencies that linear Granger causality cannot capture. Meteorological features exhibiting significant predictive improvement were selected via Random Forest and combined with influenza autocorrelation features as inputs to an LSTM-Transformer hybrid model. The model was trained using a rigorous time-series split (60%-20%-20%) and evaluated primarily using the Weighted Mean Absolute Percentage Error (WMAPE). Model explainability was assessed using SHapley Additive exPlanations (SHAP). The nonlinear Granger causality test identified significant statistical predictive relationships for all eight meteorological factors. Temperature-related variables exhibited a 6-month lag (improvement rates: 22.56%-23.99%). Precipitation showed a 1-month lag (improvement rate: 19.37%). Sunshine duration displayed a persistent predictive pattern across lags of 1 to 6 months (improvement rates: 15.8%-22.85%). The hybrid model achieved an R² of 0.786 and a WMAPE of 30.19% on the independent test set. SHAP-based explainability analysis ranked sunshine duration lagged by 2 months as the most influential predictor (mean absolute SHAP value = 0.0609), followed by the 6-month rolling mean of minimum temperature (0.0322). Meteorological factors offered important statistical predictive information for influenza incidence in Gansu Province, with sunshine duration exhibiting a persistent predictive pattern across multiple lags. The hybrid model demonstrated robust predictive performance on the independent test set.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.