A Hybrid Deep Learning-Mechanistic Modeling Framework for Dengue Transmission Dynamics in Guangdong, China

Dengue fever is a mosquito-borne viral disease with strong seasonality, periodicity, and spatial heterogeneity, posing a persistent global public health threat. We present a hybrid modeling framework that couples a recurrent neural network (RNN) with a typical Susceptible–Infected–Recovered (SIR) compartmental model to investigate dengue transmission dynamics in Guangdong, China, from 2016 to 2024. By incorporating mosquito surveillance data, meteorological variables (temperature, humidity, and precipitation), public health intervention intensity, and reported dengue cases, our model captures the complex interactions among climate, vector density, intervention policies, and disease spread. After evaluating multiple RNN architectures, the Long Short-Term Memory (LSTM) model was selected for its superior performance in predicting the mosquito ovitrap index (MOI) from climatic variables, providing a data-driven proxy for vector abundance used in the SIR model. The hybrid model is constructed upon a partially observed Markov process (POMP) and calibrated using iterated filtering for parameter optimization. Model results identify climate variables as dominant drivers of dengue transmission, while non-pharmaceutical interventions (NPIs) during the COVID-19 pandemic significantly suppressed case numbers. Scenario analysis indicates that moderate NPIs could effectively reduce outbreak magnitude, with the model estimating that approximately 46,120 dengue cases were averted during the pandemic period. This study highlights the utility of integrating deep learning with mechanistic modeling to improve understanding and forecasting of vector-borne disease dynamics. The proposed framework offers a robust and interpretable approach for developing climate- and vector-informed early warning systems and informing data-driven public health decision-making.