A Comparative Study on Deep Learning Models for COVID-19 Forecast.

Area of Science:

• Epidemiology
• Computational Biology
• Machine Learning

Background:

• The COVID-19 pandemic necessitates accurate forecasting for effective public health interventions and resource management.
• Existing mathematical and computational models face challenges due to the evolving nature of the virus.
• Deep neural networks (DNNs) offer advanced capabilities for improving prevalence predictions by integrating diverse data sources.

Purpose of the Study:

• To evaluate and compare the performance of three deep neural networks (DNNs) in modeling and forecasting COVID-19 spread.
• To investigate the impact of training data volume on prediction accuracy and long-term forecasting capabilities.
• To identify the most reliable DNN model for accurate COVID-19 dynamics forecasting.

Main Methods:

• Employed Long Short-Term Memory (LSTM) network, Physics-informed Neural Network (PINN), and Deep Operator Network (DeepONet) for COVID-19 spread modeling.
• Utilized global COVID-19 case data from 2021 (CSSE, Johns Hopkins University) for training and testing.
• Applied a seven-day moving average and normalization techniques to stabilize deep learning model training.

Main Results:

• Systematic investigation of training data effects on prediction accuracy and long-term forecast capability for each model.
• DeepONet demonstrated superior performance across all test cases compared to LSTM and PINN.
• Relative L2 errors indicated DeepONet's higher accuracy in forecasting COVID-19 dynamics.

Conclusions:

• DeepONet, by learning hidden physics from data, proves to be a reliable tool for accurate COVID-19 forecasting.
• The findings support the use of advanced machine learning models for enhanced pandemic prediction and response.
• Optimized model selection can improve the effectiveness of public health strategies and resource allocation during health crises.