EpiDeep

Influenza leads to regular losses of lives annually and requires careful monitoring and control by health organizations. Annual influenza forecasts help policymakers implement effective countermeasures to control both seasonal and pandemic outbreaks. Existing forecasting techniques suffer from problems such as poor forecasting performance, lack of modeling flexibility, data sparsity, and/or lack of intepretability. We propose EpiDeep, a novel deep neural network approach for epidemic forecasting which tackles all of these issues by learning meaningful representations of incidence curves in a continuous feature space and accurately predicting future incidences, peak intensity, peak time, and onset of the upcoming season. We present extensive experiments on forecasting ILI (influenza-like illnesses) in the United States, leveraging multiple metrics to quantify success. Our results demonstrate that EpiDeep is successful at learning meaningful embeddings and, more importantly, that these embeddings evolve as the season progresses. Furthermore, our approach outperforms non-trivial baselines by up to 40%.