Leveraging regularity in COVID-19 growth rate dynamics for epidemic wave forecasting

Area of Science:

• Epidemiology
• Computational Biology
• Public Health

Background:

• Real-time infectious disease forecasting is critical for public health.
• Traditional methods struggle with non-seasonal epidemics (e.g., COVID-19) due to irregular patterns and limited training data.
• Existing models rely on stable epidemic wave characteristics like duration, peak timing, and magnitude.

Purpose of the Study:

• To develop an improved Bayesian forecasting framework for infectious diseases lacking strict seasonality.
• To address limitations of traditional methods in predicting epidemic trajectories.
• To enhance the accuracy and reliability of short-term and medium-term epidemic forecasts.

Main Methods:

• Developed a Bayesian framework focusing on growth rate dynamics instead of incidence.
• Utilized past epidemic waves to establish priors on growth rate trajectory shapes.
• Incorporated Gaussian processes for real-time data smoothing and growth rate estimation.
• Updated forecasts dynamically as new data became available.

Main Results:

• Achieved a 27%-61% improvement in weighted interval score for 14-day ahead forecasts compared to baseline models.
• Demonstrated ability to predict medium-term statistics including peak timing and magnitude.
• Reduced root mean squared error by 41% using Gaussian processes for growth rate estimation on simulated data.

Conclusions:

• The proposed Bayesian framework offers a promising alternative for forecasting non-seasonal infectious diseases.
• The method shows significant improvements over traditional techniques, particularly when seasonality is absent.
• Further research into non-mechanistic time series models for epidemic forecasting is warranted.