Construction of an early warning model for imported dengue fever: a multi-fusion machine learning approach

Objective  To develop an imported dengue fever prediction model based on multi-fusion machine learning, aiming to enhance the scientific basis and foresight of infectious disease prevention and control at Guangzhou port.

Methods Data on imported dengue fever cases at Guangzhou port from 2015 to 2019 were collected, along with meteorological indicators, inbound passenger volume, flight traffic, vegetation index, and vector density. After data cleaning and standardization, three base models (BP neural network, random forest, and CatBoost) were constructed. Weighted averaging and stacking ensemble strategies were then adopted to build dual-fusion and triple-fusion models. Model performance was evaluated based on mean squared error (MSE) and coefficient of determination (R²).

Results The fusion models outperformed single models overall, demonstrating higher predictive accuracy and stability. Among them, the stacked ensemble of random forest and CatBoost achieved the best performance, with an MSE of 4.749 and an R² of 0.601, significantly improving upon the best single model. SHAP analysis further revealed that humidity, sunshine duration, and diurnal temperature variation were the main driving factors, while inbound passenger volume also contributed substantially to predictions.

Conclusions Multi-fusion machine learning models provide significant advantages in predicting imported dengue cases, offering strong support for early warning and the formulation of prevention strategies in port public health.