Unlocking the Predictive Power of Heterogeneous Data to Build an Operational Dengue Forecasting System

Abstract

Predicting an infectious disease can help reduce its impact by advising public health interventions and personal preventive measures. While availability of heterogeneous data streams and sensors such as satellite imagery and the Internet have increased the opportunity to indirectly measure, understand, and predict global dynamics, the data may be prohibitively large and/or require intensive data management while also requiring subject matter experts to properly exploit the data sources (e.g., deriving features from fundamentally different data sets). Few efforts have quantitatively assessed the predictive benefit of novel data streams in comparison to more traditional data sources, especially at fine spatio-temporal resolutions. We have combined multiple traditional and non-traditional data streams (satellite imagery, Internet, weather, census, and clinical surveillance data) and assessed their combined ability to predict dengue in Brazils 27 states on a weekly and yearly 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. We also predict yearly cumulative risk by municipality and state. The top-performing model and utility of predictive data varies by state, implying that forecasting and nowcasting efforts in the future may be made more robust by and benefit from the use of multiple data streams and models. One size does not fit all, particularly when considering state-level predictions as opposed to the whole country. Our first-of-its-kind high resolution flexible system for predicting dengue incidence with heterogeneous (and still sometimes sparse) data can be extended to multiple applications and regions.