Developing a dengue forecast model using Long Short Term Memory neural networks method

Background Dengue Fever (DF) is a tropical mosquito-borne disease that threatens public health and causes enormous economic burdens worldwide. In China, DF expanded from coastal region to inner land, and the incidence sharply increased in the last few years. In this study, we conduct the analysis of dengue using the Long Short Term Memory (LSTM) recurrent neural networks. This is an artificial intelligence technology, to develop a precise dengue forecast model. Methodology/Principal Findings The model is developed from monthly dengue cases and local meteorological data of 2005–2018 among top 20 Chinese cities with a record of the highest dengue incidence. The first 13 year data were used to construct the LSTM and to predict the dengue outbreaks in 2018. The results are compared with the estimated dengue cases of other previously published models. Model performance and prediction accuracy were assessed using Root Mean Square Error (RMSE). With the LSTM method, the prediction measurements of average RMSE drop by 54.79% and 34.76% as compared with the Susceptible Infected Recovered (SIR) model and Zero Inflated Generalized Additive Model (ZIGAM). Our results showed that if only local data were used to develop forecast models, the LSTM neural networks would fail to capture the transmission characteristics of dengue virus in areas with fewer dengue cases. Contrarily, transfer learning (TL) can improve the accuracy of prediction of the LSTM neural network model in areas with fewer dengue incidences. Conclusion and significance The LSTM model is beneficial in predicting dengue incidence as compared with other previously published forecasting models. The findings provide a more precise forecast dengue model, which can help the local government and health-related departments respond early to dengue epidemics.