Large-scale multivariate forecasting models for Dengue - LSTM versus random forest regression.

Area of Science:

• Epidemiology
• Machine Learning
• Public Health

Background:

• Effective management of seasonal diseases like dengue fever requires timely interventions before peak transmission seasons.
• Accurate incidence forecasts are vital for controlling dengue, but traditional time series models struggle with prediction accuracy.
• Epidemic seasons fluctuate annually, necessitating advanced forecasting methods.

Purpose of the Study:

• To compare machine learning models for forecasting weekly dengue incidence in 790 Brazilian cities.
• To evaluate the performance of feature selection methods (LASSO, Random Forest) and LSTM in dengue forecasting.
• To incorporate spatial data from similar cities to improve transmission modeling.

Main Methods:

• Proposed and compared machine learning models including LASSO, Random Forest regression, and LSTM (a deep recurrent neural network).
• Utilized multivariate time-series data as predictors for forecasting.
• Incorporated time-series data from similar cities to capture spatial disease transmission dynamics.

Main Results:

• The LSTM recurrent neural network model demonstrated the highest performance in predicting future dengue incidence.
• The model's effectiveness was validated across cities of varying sizes in Brazil.
• Machine learning approaches outperformed classical time series methods for this task.

Conclusions:

• LSTM deep recurrent neural networks offer superior accuracy for forecasting dengue incidence compared to traditional methods.
• Machine learning models integrating spatial data show promise for enhancing infectious disease surveillance and control.
• Accurate dengue forecasting using advanced computational models can significantly aid public health interventions.