Modelagem preditiva de surtos epidêmicos usando redes neurais LSTM: uma aplicação para a covid-19

Abstract

The COVID-19 pandemic, in its initial three years, resulted in a time series that was difficult to predict due to public policy actions and the emergence of new variants. Mobility restrictions, new variants, vaccines, and cultural differences led to distinct stages of virus spread, making it challenging for a single epidemiological prediction model to forecast the entire time series, thus necessitating different models. Based on the application of neural networks in epidemiological studies for time series prediction, this work aims to develop a methodology based on a multi-layer Long Short-Term Memory (LSTM) neural network, capable of being applied to nearly three years of pandemic data to predict the number of new daily cases of the disease. The methodology applies a series of tests, combining different network input data, such as daily cases and vaccinations to predict the number of cases for several days into the future. The experiments were conducted for Brazil and other countries, showing good results for predictions up to fifteen days in advance, with the potential to identify trend changes in the daily case timeline. This capability is useful for detecting the onset of new waves of infection, contributing to public health alert systems and immediate decision-making.