Machine Learning for Dengue Outbreak Prediction: An AI-Driven Public Health Early Warning System

Introducing the PredictIA_Dengue open-source project, which uses machine learning to integrate multi-source data including historical outbreaks, climate, environment, and socioeconomic factors to build predictive models that identify dengue outbreak risks in advance. This helps public health departments break through the limitations of traditional passive monitoring, deploy prevention and control measures (such as mosquito vector control, resource stockpiling, and public education) early, and protect community health.

Dengue is an acute mosquito-borne infectious disease caused by the dengue virus transmitted via Aedes mosquitoes. The WHO estimates 390 million global infections annually (96 million symptomatic cases, 20,000 deaths). Its harms include direct health impacts, pressure on healthcare systems (hospital overcrowding during epidemic seasons), and socioeconomic burdens (especially in developing countries). Climate change, urbanization, and globalization have expanded its spread. Traditional passive monitoring (responding only after cases appear) has issues like time lag, resource waste, and missing the golden window for prevention and control, necessitating proactive intelligent early warning methods.

Data Sources: Integrates data from historical outbreaks (confirmed cases/hospitalizations/deaths), climate (temperature/rainfall/humidity), environment (vegetation/water bodies/urbanization via satellite remote sensing), socioeconomic factors (population density/poverty rate/sanitation facilities), and vector monitoring (mosquito density/virus infection rate). Feature Engineering: Time series processing (sliding window), spatial features (spatial lag), lag features (multi-period delay effects), interaction features (multi-dimensional interactions like climate-environment). Model Selection: Baseline models (ARIMA/SARIMA), machine learning models (Random Forest/XGBoost/LightGBM), deep learning models (LSTM/attention mechanisms), and ensemble strategies (average/weighted/stacked) to enhance robustness.

Time Scale: Short-term (1-4 weeks for immediate resource allocation), medium-term (1-3 months for seasonal planning), long-term (6-12 months for strategic planning). Spatial Scale: Regional level (city/province for decision support), local level (community/neighborhood for targeted intervention), hotspot identification (priority deployment of preventive measures).

Data Quality: Reduce the impact of reporting delays, underreporting, and diagnostic inconsistencies via data cleaning and imputation, and robust models; Imbalanced Data: Handle class imbalance caused by rare outbreaks using resampling and cost-sensitive learning; Concept Drift: Establish model monitoring and update mechanisms to address changes in transmission patterns; Interpretability: Use feature importance and SHAP values to provide basis for predictions and enhance decision-makers' trust.

The prediction system has been applied to:

• Public health decision-making: data-driven resource allocation and targeted strategy formulation;
• Healthcare resource planning: pre-preparing beds, medicines, and medical staff;
• Mosquito vector control: identifying high-risk areas for priority mosquito eradication;
• Public communication: designing educational activities to remind people of protection measures.

Scalability: The methodology can be transferred to other mosquito-borne diseases like Zika and Chikungunya; Future Directions: Multi-disease integrated early warning, real-time monitoring and automatic alerts, policy impact assessment (simulating intervention scenarios), and global collaboration networks (cross-border data sharing and response).

Privacy Protection: Use data desensitization and differential privacy to protect personal health information; Fairness: Fairness audits to ensure no systemic bias in the model; Transparency and Accountability: Document model assumptions, limitations, and uncertainties; Human-Machine Collaboration: AI assists rather than replaces expert judgment, combining model capabilities with expert experience.