[Paper Review] Time Series Methods and Ensemble Models to Nowcast Dengue at the State Level in Brazil

Katherine Kempfert, Kaitlyn Martinez+9 more|arXiv (Cornell University)|Jun 3, 2020

Mosquito-borne diseases and control72 references4 citations

TL;DR

This study develops and evaluates time series and ensemble models to nowcast dengue incidence at the state level in Brazil using integrated data streams—clinical surveillance, weather, satellite imagery, and Internet search trends. It achieves high accuracy, with 25 states exceeding 80% correlation and a median correlation of 91.75% across 27 states, demonstrating that exogenous data significantly improves predictions and that ensemble methods enhance robustness and reliability for real-time public health decision-making.

ABSTRACT

Predicting an infectious disease can help reduce its impact by advising public health interventions and personal preventive measures. Novel data streams, such as Internet and social media data, have recently been reported to benefit infectious disease prediction. As a case study of dengue in Brazil, we have combined multiple traditional and non-traditional, heterogeneous data streams (satellite imagery, Internet, weather, and clinical surveillance data) across its 27 states on a weekly basis over seven years. For each state, we nowcast dengue based on several time series models, which vary in complexity and inclusion of exogenous data. The top-performing model varies by state, motivating our consideration of ensemble approaches to automatically combine these models for better outcomes at the state level. Model comparisons suggest that predictions often improve with the addition of exogenous data, although similar performance can be attained by including only one exogenous data stream (either weather data or the novel satellite data) rather than combining all of them. Our results demonstrate that Brazil can be nowcasted at the state level with high accuracy and confidence, inform the utility of each individual data stream, and reveal potential geographic contributors to predictive performance. Our work can be extended to other spatial levels of Brazil, vector-borne diseases, and countries, so that the spread of infectious disease can be more effectively curbed.

Motivation & Objective

• To improve real-time dengue nowcasting in Brazil by integrating diverse data streams including clinical surveillance, weather, satellite imagery, and Internet search trends.
• To evaluate the predictive contribution of novel data streams (e.g., satellite and Internet data) relative to traditional clinical surveillance and weather data.
• To develop and validate ensemble modeling approaches that combine multiple time series models to enhance predictive robustness and accuracy at the state level.
• To identify geographic and socioeconomic factors associated with model performance variation across Brazilian states.
• To provide a transferable framework for nowcasting other vector-borne diseases in Brazil and similar countries using heterogeneous data sources and ensemble modeling techniques.

Proposed method

• Combines weekly dengue case counts from Brazil’s Ministry of Health with exogenous data: satellite-derived vegetation and temperature indices, weather data, and Internet search query volumes.
• Applies four time series models per state: SARIMA, VAR, STL decomposition with local regression (LOESS), and their variants, each incorporating different subsets of exogenous variables.
• Uses a trimmed mean ensemble approach to combine predictions from individual models, removing the top and bottom 20% of forecasts to reduce outlier influence.
• Employs a weighted mean ensemble using model weights derived from validation performance—weights are proportional to the frequency with which each model minimizes absolute error on the validation set.
• Computes 95% prediction intervals conservatively by taking the min and max of the prediction interval bounds from the trimmed set of models.
• Validates models on a 2015–2016 testing window, using Pearson correlation and empirical coverage of prediction intervals as primary evaluation metrics.

Experimental results

Research questions

• RQ1 How do different combinations of traditional and non-traditional data streams (e.g., satellite, weather, Internet) affect dengue nowcasting accuracy at the state level in Brazil?
• RQ2 Which time series models perform best for dengue nowcasting across different Brazilian states, and does the optimal model vary spatially?
• RQ3 To what extent do ensemble models improve predictive performance and reliability compared to individual models in real-time dengue nowcasting?
• RQ4 What are the key socioeconomic and geographic factors associated with variations in model performance across Brazilian states?
• RQ5 Can a single, unified ensemble framework effectively generalize across diverse epidemiological and environmental conditions in Brazil’s 27 states?

Key findings

• The trimmed mean ensemble achieved a median Pearson correlation of 91.75% between fitted and observed dengue cases across all 27 Brazilian states during the 2015–2016 testing window.
• 25 out of 27 states achieved a Pearson correlation coefficient above 80% using the trimmed mean ensemble, with the highest correlation reaching 96.44% in one state.
• The 95% prediction intervals achieved empirical coverage of at least 96% for half of the states, indicating reliable uncertainty quantification.
• Incorporating exogenous data—particularly weather or satellite-derived environmental variables—improved model performance, and including just one such stream often yielded performance comparable to using all data streams.
• Model performance showed significant spatial autocorrelation and was positively associated with state-level indicators of education, employment, and population size.
• The optimal individual model varied by state, justifying the use of ensemble methods to achieve robust, state-specific nowcasting without prior knowledge of the best-performing model per region.