A novel bidirectional LSTM deep learning approach for COVID-19 forecasting.

Area of Science:

• Epidemiology
• Data Science
• Computational Biology

Background:

• COVID-19 caused significant global morbidity and mortality.
• Accurate forecasting of COVID-19 cases is crucial for public health response.
• Early pandemic stages presented unique challenges for predictive modeling.

Purpose of the Study:

• To develop and validate a deep-learning model for forecasting daily COVID-19 cases.
• To assess the model's performance in the early stages of the pandemic across 190 countries.
• To compare the deep-learning model's accuracy against a classical ARIMA model.

Main Methods:

• Utilized a Bidirectional Long-Short Term Memory (Bi-LSTM) deep-learning architecture.
• Trained models on daily confirmed cases, reproduction number, policy measures, mobility, and flight data from January 2020 to January 2021.
• Forecasted new daily COVID-19 cases 14 days in advance for 190 countries.

Main Results:

• The Bi-LSTM models demonstrated comparable accuracy to each other and outperformed the ARIMA model in total absolute percentage error.
• Median Mean Absolute Error (MAE) was 157 and 150 for the two Bi-LSTM models, respectively.
• Countries with higher case numbers and more infection waves generally had more accurate forecasts.

Conclusions:

• A deep-learning approach using Bi-LSTM architecture effectively forecasts COVID-19 cases in the early pandemic.
• Open-source data can be leveraged for robust epidemiological predictions.
• Model accuracy may be maintained with a reduced set of input variables.