Machine Learning-Driven COVID-19 Hospitalization Forecasting: From Theory to Practice in a Major Northeastern...

Area of Science:

• Epidemiology
• Machine Learning
• Public Health

Background:

• Accurate prediction of respiratory virus waves, including COVID-19, is vital for public health.
• Existing COVID-19 forecasting models require performance enhancements.

Purpose of the Study:

• To develop and evaluate an advanced machine learning model for forecasting COVID-19 hospitalizations.
• To improve the accuracy and utility of predictive models for pandemic response.

Main Methods:

• Extended the Neural Basis Expansion Analysis for Time Series Forecasting (N-BEATS) architecture.
• Integrated temporal convolutional networks for exogenous variables and residual blocks for probabilistic predictions.
• Compared performance against COVID-19 Forecast Hub ensembles and implemented transfer learning in a hospital setting.

Main Results:

• Achieved a 34.0% improvement in mean absolute error (MAE) over the weighted ensemble for US COVID-19 hospitalizations.
• Demonstrated superior performance using mean absolute percent error (MAPE) and symmetric mean absolute percent error (sMAPE).
• Provided actionable forecasts for hospital resource allocation and surge planning.

Conclusions:

• The enhanced model significantly improves COVID-19 hospitalization forecasting, especially for peaks and resurgences.
• Successful real-world implementation demonstrates potential for aiding decision-making during respiratory virus outbreaks.
• Highlights the value of advanced machine learning in pandemic preparedness and resource management.