Forecasting Covid-19 Dynamics in Brazil: A Data Driven Approach.

Area of Science:

• Epidemiology
• Computational Biology
• Artificial Intelligence

Background:

• The COVID-19 pandemic presented unprecedented challenges for public health surveillance and prediction.
• Accurate forecasting models are crucial for resource allocation and policy-making during pandemics.
• Early prediction models faced limitations, necessitating the development of more robust approaches.

Purpose of the Study:

• To introduce a novel data-driven methodology for predicting COVID-19 pandemic dynamics.
• To apply and evaluate this approach for Brazilian states, providing predictions from May 2020.
• To estimate key pandemic statistics, including infection peaks and total case numbers.

Main Methods:

• Utilized Long Short-Term Memory for Data Training-SAE (LSTM-SAE) as a baseline, followed by Modified Auto-Encoder (MAE) networks.
• Employed clustering of global regions based on pandemic response features to select optimal training data.
• Applied curve fitting to refine MAE output and estimate pandemic peak distributions.

Main Results:

• MAE models, trained on clustered global data, predicted over one million total COVID-19 infections in Brazil.
• São Paulo state was predicted to have approximately 150,000 confirmed cases.
• Most Brazilian states were projected to reach peak infections in the latter half of May 2020.

Conclusions:

• The data-driven MAE approach provides valuable predictions for COVID-19 dynamics in Brazil.
• Pandemic resolution, defined as 97% of cases reaching an outcome, was estimated between June and August 2020 across states.
• The study highlights the ongoing growth of the pandemic in Brazil during the prediction period.