Related Experiment Video
Updated: May 4, 2026
Visualizing Dengue Virus through Alexa Fluor Labeling
Published on: July 9, 2011
A novel model in predicting dengue fever considering spatial heterogeneity of nine impact factors and four AI
Tianxiang Jiang1, Qianqian Zhang2, Mengxuan Chen1
1School of Global Health, Chinese Center for Tropical Diseases Research, Shanghai Jiao Tong University School of Medicine, Shanghai 200025, China; School of Public Health, Shanghai Jiao Tong University School of Medicine, Shanghai 200025, China.
Background:
Previous studies have explored the relationships between dengue fever (DF) and its impact factors, using various models. However, few have considered the spatial heterogeneity of these relationships in simulating DF variations.
Methods:
This study analyzed monthly DF incidence and rate data across 34 provincial-level administrative divisions (PLADs) in China from 2004 to 2019, along with climatic and socioeconomic variables. Nine impact factors were included: six climatic variables - minimum temperature (Tmin), mean temperature (Tmean), maximum temperature (Tmax), relative humidity (RH), precipitation (PRCP), and the El Niño-Southern Oscillation (ENSO) - and three socioeconomic variables - population density (PD), per capita regional gross domestic product (pcGDP) and number of foreign visitor arrivals (NFVAs). We introduced a modeling system that incorporates the spatial heterogeneity of these factors and integrates four artificial intelligence (AI) algorithms: support vector machine (SVM), artificial neural network (ANN), random forest (RF), and gradient boosting machine (GBM). This system was implemented under a novel framework - grid-by-grid, multi-algorithms, optimal combination (GGMAOC) - to explore the spatial heterogeneity of both algorithms and impact factors.
Results:
From 2004 to 2019, DF in China showed a significant upward trend, an average annual increase of 1618 cases. Among the impact factors, socioeconomic variables exhibited stronger associations, with correlation coefficients exceeding 0.77. In Yunnan, the ANN model performed best (DISO=0.18), with PDLag3 contributing over 31%. In Guangdong, the RF model was optimal (DISO=0.26), with TmaxLag2 contributing over 67%. Performance varied across PLADs, which highlights the spatial heterogeneity of both impact factors and algorithmic responses. Importantly, the proposed GGMAOC framework substantially outperformed the traditional stepwise regression approach.
Conclusion:
Our findings reveal that impact factor patterns vary across PLADs, emphasizing spatial heterogeneity and the need for PLAD-specific modeling. The GGMAOC framework identifies key impact factors and optimal models for DF dynamics, offering high predictive confidence and broad applicability to other diseases with spatially heterogeneous impact factors. Furthermore, the successful model construction in GD and YN and the identification of challenges in TW modeling further support the necessity and innovation of adopting a spatially heterogeneous modeling framework in infectious disease modeling as proposed in this study.
