Combining weather patterns and cycles of population susceptibility to forecast dengue fever epidemic years in Brazil: a dynamic, ensemble learning approach

Sarah F. McGough, César Leonardo Clemente, J. Nathan Kutz, Mauricio SantillanabioRxiv 2019 · DOI

TL;DR

This paper introduces a novel machine learning approach to forecast dengue fever outbreaks in Brazil months in advance, by dynamically learning from weather patterns and incorporating population susceptibility cycles. This method significantly improves prediction accuracy compared to weather-only models, offering a valuable tool for public health preparedness.

What Problem Does This Paper Address?

Dengue fever is a major global health concern, with transmission driven by a complex interplay of human behavior, climate, and mosquito populations. Accurately forecasting dengue outbreaks is crucial for effective resource allocation and public health interventions, but existing methods often struggle with the dynamic and spatially heterogeneous nature of the disease. Specifically, the authors highlight the lack of a systematic, adaptive, and generalizable framework that can identify predictive patterns in weather and population susceptibility at a city level, especially months before an outbreak. They also point out that many models neglect the crucial role of population immunity, which influences susceptibility and can lead to cyclical outbreak patterns.

Methods

The core of the paper is a data-driven, machine learning framework designed for dynamic and adaptive forecasting. The approach involves several key steps:

  • Signal Preprocessing: Identifying relevant time windows (10-95 days) within the last 7 months of the year, using a windowing technique to capture signals within and around these intervals.
  • Time Series Feature Extraction: Extracting simple summary features from two key weather variables: daily mean temperature and mean precipitation frequency.
  • Independent Model Training: Training a large number of Support Vector Machine (SVM) classifiers, each for a specific time interval and weather variable combination.
  • Model Selection: Selecting the top 11 best-performing models based on their historical accuracy and the performance of neighboring time intervals.
  • Ensemble Prediction: Combining the predictions of these top models using a majority vote to produce a final forecast.

A significant innovation is the post-hoc incorporation of empirically observed dengue susceptibility cycles (3-4 year outbreak patterns) using a decision rule based on Markov transition probabilities. This rule acts as an "expert opinion" to override weather-based predictions when historical susceptibility patterns suggest a different outcome.

Datasets

  • Dengue Fever Case Reports: Annual data for 20 dengue-endemic municipalities in Brazil from 2001-2017, obtained from the Brazilian Ministry of Health DataSUS.
  • Meteorological Data: Daily temperature (Kelvin) and hourly precipitation (kg/m²) data for the same period (2000-2016) from the GMAO-NASA MERRA-2 dataset.

The study uses a dataset of 17 years of outcome data for training and out-of-sample predictions, with a focus on a 7-year initial training period to maximize out-of-sample evaluation.

Key Insights and Ideas

  • Population Susceptibility is Key: Incorporating empirically observed dengue susceptibility cycles (proxied by 3-4 year outbreak patterns) significantly improves prediction accuracy, particularly for non-epidemic years. This highlights that immunity plays a critical, often overlooked, role in dengue dynamics.
  • Dynamic and Adaptive Learning: The framework's ability to dynamically identify predictive "weather signatures" (time windows with strong predictive power) and adapt to new information makes it robust across diverse locations and changing conditions.
  • Temporal Patterns Matter: The success of the model is linked to the identification of consistent and robust "weather signatures" in cities with higher prediction accuracy, suggesting a biological basis for these temporal patterns.
  • "Expert Opinion" for Overrides: The decision rule based on dengue cycles acts as a valuable "expert opinion" to correct potentially misleading weather-based forecasts, especially in situations where historical immunity patterns are strong indicators.

Key Results

  • The weather-based approach alone achieved an 81% hit rate (sensitivity) for epidemic years and an overall accuracy of approximately 72%.
  • Incorporating dengue cycles improved the non-epidemic detection rate (specificity) to 71% and increased the overall accuracy to 74.2%.
  • The combined approach correctly forecasted 100% of 13 outbreaks in 2016.
  • The model demonstrated success in predicting outbreaks months in advance, providing timely information for public health decision-making.
  • The study identified distinct "weather signatures" in different regions of Brazil, suggesting that weather's influence on dengue transmission can vary geographically and seasonally.

Limitations and Future Work

  • Small Data Setting: The study acknowledges that predicting in a "small data" setting (17 outcome data points) limits the ability to establish strong climatic distinctions and stable identification of susceptibility cycles. Longer time series data would likely improve performance.
  • Limited Predictors: The current model primarily uses temperature and precipitation. The authors suggest incorporating a wider range of environmental variables (e.g., humidity, vegetation, soil moisture) and non-environmental factors (e.g., human activity, public health interventions) could further enhance accuracy.
  • Serotype Information: The paper notes that incorporating information on specific circulating DENV serotypes could improve predictions, especially in understanding why certain years with predicted low susceptibility might still experience outbreaks.
  • Generalizability: While the framework is designed to be generalizable, its performance in other regions would require validation. The authors highlight that the framework can be extended to accommodate additional predictors and longer time series for improved performance in different settings.

Summary updated

Abstract

Abstract Transmission of dengue fever depends on a complex interplay of human, climate, and mosquito dynamics, which often change in time and space. It is well known that disease dynamics are highly influenced by a population’s susceptibility to infection and microclimates, small-area climatic conditions which create environments favorable for the breeding and survival of the mosquito vector. Here, we present a novel machine learning dengue forecasting approach, which, dynamically in time and adaptively in space, identifies local patterns in weather and population susceptibility to make epidemic predictions at the city-level in Brazil, months ahead of the occurrence of disease outbreaks. Weather-based predictions are improved when information on population susceptibility is incorporated, indicating that immunity is an important predictor neglected by most dengue forecast models. Given the generalizability of our methodology, it may prove valuable for public-health decision making aimed at mitigating the effects of seasonal dengue outbreaks in locations globally.
BibTeX
@misc{mcgough2019combining,
  author = {Sarah F. McGough and César Leonardo Clemente and J. Nathan Kutz and Mauricio Santillana},
  title = {{Combining weather patterns and cycles of population susceptibility to forecast dengue fever epidemic years in Brazil: a dynamic, ensemble learning approach}},
  year = {2019},
  doi = {10.1101/666628},
  howpublished = {bioRxiv},
}

Loading related content...

Figure