Assessing dengue forecasting methods: a comparative study of statistical models and machine learning techniques in...

Area of Science:

• Epidemiology
• Public Health
• Computational Science

Background:

• Dengue is a significant mosquito-borne viral disease threat in tropical and subtropical regions.
• Accurate dengue outbreak forecasting is vital for public health planning and interventions.
• This study evaluates predictive models for dengue forecasting to improve surveillance systems.

Purpose of the Study:

• To assess the predictive performance and computational efficiency of statistical and machine learning models for dengue forecasting.
• To compare models with and without climate factors (temperature, humidity).
• To inform the design of effective dengue surveillance systems.

Main Methods:

• Compared statistical models (ARIMA, SARIMAX) and machine learning techniques (LSTM, Prophet, Random Forest, XGBoost, SVM).
• Utilized a dynamic window approach for weekly forecasts at multiple time horizons.
• Incorporated climate covariates and lagged climatic variables to account for transmission dynamics.

Main Results:

• SARIMAX improved accuracy over ARIMA by including climate covariates.
• LSTM with climate covariates was the most accurate machine learning model, though slower.
• Prophet with climate covariates excelled in long-term forecasts; ensemble models showed significant improvements.

Conclusions:

• Various dengue forecasting methods have distinct strengths and limitations across timeframes.
• Machine learning techniques and climate covariate integration significantly enhance forecast accuracy.
• Findings support public health officials in developing better dengue surveillance and resource allocation strategies.