Improving outbreak forecasts through model augmentation

Area of Science:

• Epidemiology
• Computational Biology
• Public Health

Background:

• Accurate disease outbreak forecasts are vital for public health preparedness and resource allocation.
• Existing forecasting models (empirical and mechanistic) often falter during rapid epidemic escalation.
• There is a need for improved prediction accuracy during critical outbreak periods.

Purpose of the Study:

• To introduce epimodulation, a novel hybrid approach to enhance disease outbreak forecasting.
• To integrate fundamental epidemiological principles into existing predictive models.
• To improve forecast accuracy, particularly around epidemic peaks.

Main Methods:

• Developed and applied the epimodulation technique.
• Integrated epimodulation with various empirical and machine learning models (ARIMA, Holt-Winters, GBM, Prophet, Spline).
• Evaluated performance on COVID-19 and influenza hospital admission data, including complex ensemble models.

Main Results:

• Epimodulation improved overall prediction accuracy by 12.3% for COVID-19 and 32.9% for influenza hospital admissions.
• Accuracy during epidemic peaks saw substantial improvements: 27.9% for COVID-19 and 43.8% for influenza.
• Enhanced the performance of complex models like the COVID-19 Forecast Hub ensemble.

Conclusions:

• Epimodulation offers a broadly applicable method to significantly boost disease forecasting reliability.
• The hybrid approach improves predictions, especially during critical epidemic escalation and peak phases.
• This enhances preparedness for public health emergencies through more accurate outbreak predictions.