An In-depth Analysis on the Use of Long Short-term Memory Networks to Predict Incidence and Prevalence of Covid-19 in Latin America

Abstract

The use of machine learning techniques, especially deep learning, could improve the predictions of the currently used epidemiological models for predicting Covid-19 in the short term. This information is essential for better decision making and to reduce the impacts of the disease spread in different countries. We explored the use of support vector regression (SVR) and long short-term memory networks (LSTM), the state of the art neural network architecture for time series analysis, to predict the daily incidence and prevalence for nine countries in Latin America. Our methodology and the models used can be replicated in other countries. Our main findings were: (i) there is no single best model or best hyperparameters configuration for all countries and targets; (ii) the LSTM showed an average MAE that was around 50% lower for incidence and 20% lower for prevalence when considering all countries; (iii) the LSTM showed better results for predicting incidence for most countries (Argentina, Bolivia, Brazil, Guatemala, and Haiti); (iv) the SVR showed better results for predicting prevalence for most countries (Argentina, Bolivia, Colombia, Cuba, Guatemala, and Haiti); and (v) for Brazil, the LSTM provided better results for both targets, with an MAE that was 68% lower for incidence and 73% lower for prevalence.