Dengue fever prediction based on meteorological features and deep learning models

Area of Science:

• Epidemiology
• Environmental Science
• Data Science

Background:

• Dengue fever epidemics pose a significant global health challenge, necessitating accurate prediction of epidemiological trends.
• Meteorological factors like temperature, humidity, and precipitation are known to influence dengue fever's occurrence and prevalence.
• Existing prediction models may not fully capture the complex interplay of multidimensional meteorological features.

Purpose of the Study:

• To propose an effective hybrid model for improving dengue fever prediction performance.
• To incorporate multidimensional meteorological features into dengue trend forecasting.
• To address data scarcity issues in epidemiological datasets.

Main Methods:

• Data augmentation using Time-series Generative Adversarial Networks (TimeGAN).
• Meteorological data decomposition via Symplectic Geometry Mode Decomposition (SGMD) and reconstruction using Sample Entropy (SE).
• Feature extraction and fusion using bidirectional temporal convolutional networks (BiTCN) and attention-based bidirectional long and short-term memory networks (BiLSTM).

Main Results:

• The proposed hybrid model demonstrated accurate prediction of dengue fever epidemic trends in Guangdong Province, China.
• Achieved a Mean Absolute Error (MAE) of 192.98759 and a Mean Absolute Percentage Error (MAPE) of 2.492.
• The integration of TimeGAN, SGMD, SE, BiTCN, and BiLSTM significantly improved prediction accuracy.

Conclusions:

• The developed hybrid model offers a robust approach for predicting dengue fever epidemics.
• Accurate forecasting of dengue trends is vital for effective public health interventions and resource allocation.
• This methodology highlights the potential of advanced machine learning techniques in analyzing complex environmental and epidemiological data.