Machine Learning and Probabilistic Approaches for Forecasting COVID-19 Transmission and Cases

Area of Science:

• Epidemiology
• Machine Learning
• Biostatistics

Background:

• Accurate forecasting of COVID-19 transmission is crucial for public health.
• Existing methods like EpiNow2 provide valuable estimates but can be improved.

Purpose of the Study:

• To develop and evaluate a machine learning framework for predicting the effective reproductive number (Rt) and COVID-19 case counts.
• To enhance forecasting accuracy and robustness at the county level in South Carolina.

Main Methods:

• Developed a probabilistic forecasting framework using machine learning models (regression, Random Forest, XGBoost).
• Integrated initial Rt estimates from EpiNow2 with spatial smoothing.
• Utilized a probabilistic Poisson model for case count predictions.
• Employed an ensemble approach combining multiple models.

Main Results:

• The ensemble model consistently outperformed EpiNow2 in forecasting Rt and case counts across 7, 14, and 21-day horizons.
• Achieved a median percentage agreement (PA) of 94.4% for 7-day Rt forecasts in the first period, compared to 87.0% for EpiNow2.
• Demonstrated improved stability and performance in case count forecasting.

Conclusions:

• Combining spatial smoothing with ensemble machine learning models significantly enhances epidemic forecasting accuracy and robustness.
• The developed framework offers a more reliable tool for public health decision-making regarding COVID-19.