Dengue fever prediction based on meteorological features and deep learning models

The dengue fever epidemic is one of the health priorities of the World Health Organization (WHO), and accurately predicting its epidemiological trends is crucial. Multi source geographic data such as temperature, humidity, and precipitation affect the occurrence and prevalence of dengue fever. Therefore, an effective hybrid model to improve the prediction performance of dengue fever is proposed considering the effects of multidimensional meteorological features. Initially, to address the issue of data scarcity, the Time-series Generative Adversarial Networks (TimeGAN) algorithm is employed to expand the dengue dataset. Second, the meteorological series are decomposed by Symplectic Geometry Mode Decomposition (SGMD), and then the sub-sequences are reconstructed into high, mid and low frequency signals by utilizing Sample Entropy (SE). Subsequently, a sliding window technique is applied to the signal band to accurately capture the key time periods affecting dengue fever. Finally, bidirectional temporal features related to the dengue sequence are extracted and fused by a bidirectional temporal convolutional network (BiTCN), and the fused features are inputted into a bidirectional long and short-term memory network (BiLSTM) introduced the attention module for prediction, and to obtain the final prediction results. Using the data of dengue fever cases in Guangdong Province, China, as an example, the experimental results show that the developed method can accurately predict the trend of dengue fever epidemic, with mean absolute error (MAE) and mean absolute percentage error (MAPE) of 192.98759 and 2.492, respectively.