Predicting dengue incidence leveraging internet-based data sources. A case study in 20 cities in Brazil

Area of Science:

• Epidemiology
• Public Health
• Data Science

Background:

• Dengue outbreaks cause significant global health and economic burdens.
• Current dengue control relies heavily on vector surveillance, necessitating accurate incidence prediction.
• Existing prediction models often lack city-level granularity and a clear understanding of data source contributions.

Purpose of the Study:

• To develop and evaluate a methodological framework for city-level dengue incidence estimation.
• To compare the predictive performance of different data sources and models.
• To assess the impact of data availability delays on prediction accuracy.

Main Methods:

• Developed a framework to assess and compare dengue incidence estimates at the city level.
• Evaluated random forest and LASSO regression models using historical incidence data, weather variables, and internet search trends.
• Analyzed model performance across 20 Brazilian cities with varying characteristics.

Main Results:

• Both random forest and LASSO models effectively utilized multiple data sources for accurate dengue predictions.
• Random forest models demonstrated greater robustness by producing fewer extreme outliers.
• Internet search data were the strongest predictors for long-delay (6-8 weeks) real-time predictions.
• Seasonal and short-term autocorrelation were dominant predictors for short-delay (1-3 weeks) predictions, halving the error rate.

Conclusions:

• The developed framework provides meaningful and actionable city-level dengue incidence estimates.
• The choice of dominant data source for prediction is dependent on the delay in epidemiological data availability.
• This approach supports targeted public health interventions at a fine spatial resolution.