Machine learning and probabilistic approaches for forecasting infectious disease transmission and cases

Area of Science:

• Epidemiology and Public Health
• Computational Biology
• Machine Learning in Healthcare

Background:

• Accurate forecasting of infectious disease dynamics, including the effective reproductive number (R_t) and daily case counts, is essential for effective public health interventions.
• Existing methods for forecasting infectious diseases may have limitations in accuracy and flexibility across different diseases and geographical regions.

Purpose of the Study:

• To develop and evaluate a novel machine learning and probabilistic forecasting framework for predicting R_t and daily COVID-19 cases in South Carolina counties.
• To assess the framework's performance against established methods and its potential for generalization to other infectious diseases.

Main Methods:

• Utilized EpiNow2 for initial R_t estimation via Bayesian time-series modeling, accounting for reporting delays and incubation periods.
• Refined R_t estimates using spatial covariate-adjusted smoothing with Integrated Nested Laplace Approximation (INLA).
• Employed an ensemble of machine learning models (linear regression, random forest, XGBoost) for R_t forecasting and a Poisson model for daily case forecasting, linking R_t trajectories with historical data.

Main Results:

• The ensemble forecasting approach demonstrated superior performance compared to EpiNow2 across various forecast horizons (7, 14, and 21 days).
• Achieved a median prediction accuracy (PA) of 96.5% for 7-day R_t forecasts in the first period and 93.0% in the second period, significantly outperforming EpiNow2's 87.0% and 86.8%, respectively.
• Observed similar performance improvements for daily case count forecasts, indicating the ensemble model's enhanced predictive capabilities.

Conclusions:

• The developed framework effectively integrates Bayesian estimation, spatial smoothing, and ensemble machine learning to enhance the accuracy of infectious disease transmission and case forecasts.
• This approach provides a flexible and scalable tool that improves epidemic forecasting performance, supporting data-driven public health preparedness and response strategies.
• The methodology shows promise for application to various infectious diseases beyond COVID-19.