BeCaked: An Explainable Artificial Intelligence Model for COVID-19 Forecasting.

Area of Science:

• Epidemiology
• Artificial Intelligence
• Computational Biology

Background:

• The COVID-19 pandemic presents significant global health and economic challenges.
• Accurate forecasting of pandemic trends is crucial for effective policy-making.
• Existing deep learning models for forecasting lack explainability, hindering proactive interventions.

Purpose of the Study:

• To develop a novel forecasting model for COVID-19 that combines predictive accuracy with explainability.
• To address the limitations of current deep learning models in providing interpretable results for pandemic management.

Main Methods:

• A hybrid model, BeCaked, integrating the Susceptible-Infectious-Recovered-Deceased (SIRD) compartmental model with a Variational Autoencoder (VAE) neural network was developed.
• The model was trained and validated using global COVID-19 data from Johns Hopkins University.
• Performance was evaluated based on forecasting accuracy (R-squared) and Mean Absolute Percentage Error (MAPE).

Main Results:

• BeCaked achieved high accuracy, with R-squared values up to 0.99 and a Mean Absolute Percentage Error (MAPE) as low as 0.0026 at the country level for short-term forecasts.
• The model demonstrated strong performance at the world level, achieving an R-squared of 0.98 and 0.012 MAPE for 31-step forecasts.
• The integrated SIRD model parameters provided justifications for the BeCaked model's predictions.

Conclusions:

• The BeCaked model offers a powerful tool for accurate and explainable COVID-19 forecasting.
• Its ability to provide justifications enhances its utility for public health authorities and medical experts.
• BeCaked can support timely and informed decision-making during pandemic outbreaks.