Predicting dengue incidence in China using SARIMA and hybrid neural network models: insights for public health planning

Abstract Dengue fever remains a persistent public health challenge in tropical and subtropical regions of China, disproportionately affecting economically disadvantaged and resource-limited areas. Outbreaks of dengue can strain healthcare systems and hinder socio-economic development. Developing timely and accurate forecasting models for dengue incidence is essential to strengthen early warning systems and inform the strategic allocation of public health resources. Monthly dengue case data from 2004 to 2024 were obtained from the Chinese Center for Disease Control and Prevention. The dataset was split into training and testing subsets to develop and validate predictive models. Six models were constructed: the Seasonal autoregressive integrated moving average (SARIMA) model, Prophet model, Extreme Gradient Boosting (XGBoost) model, Long Short-Term Memory (LSTM) model, a hybrid SARIMA-back propagation neural network (SARIMA-BPNN) model, and a hybrid SARIMA-Elman recurrent neural network (SARIMA-ERNN) model. Model performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), corrected mean absolute percentage error (cMAPE), and the coefficient of determination (R 2 ). From 2004 to 2024, the annual average dengue incidence in mainland China was 0.486 cases per 1,000,000 people, ranging from 0.031 to 34.161 per 1,000,000. All six models captured the general temporal trends, while the hybrid SARIMA-BPNN and SARIMA-ERNN models better reflected seasonal fluctuations and autocorrelation than the conventional model. Among them, the SARIMA-BPNN model achieved the highest predictive accuracy, particularly in capturing peak incidences and abrupt temporal changes. These findings indicate that hybrid models, especially SARIMA-BPNN, can effectively capture the nonlinear and dynamic transmission patterns of dengue. Hybrid time series models combining SARIMA with neural networks, particularly SARIMA-BPNN, provide reliable simulations of dengue incidence and can serve as valuable tools for early warning and resource allocation in high-risk and resource-limited areas. Integrating these models into public health planning may enhance epidemic preparedness and inform targeted interventions, though caution is needed when generalizing to other settings.