PREPARE: PREdicting PAndemic's REcurring Waves Amidst Mutations, Vaccination, and Lockdowns

Abstract

This study releases an adaptable framework that can provide insights to policymakers to predict the complex recurring waves of the pandemic in the medium postemergence of the virus spread, a phase marked by rapidly changing factors like virus mutations, lockdowns, and vaccinations, offering a way to forecast infection trends and stay ahead of future outbreaks even amidst uncertainty. The proposed model is validated on data from COVID-19 spread in Germany.

Figures (6)

• Figure 1: Proposed model estimated values for $I_t$ and $p$ for 9/1/20 - 1/21/21 period.(\ref{['et_i_pred']}): The observed number of infected individuals is represented by black dots, while the median estimates and future forecasts are shown in purple and orange, respectively. The shaded regions indicate the $97.5^{th}$ percentile intervals for median estimates and future forecasts. The model accurately predicts infection trends, closely matching observed data until early December, with forecasts showing a rise in infections due to mutations before vaccination. A significant forecast increase between December $1^{st}$ and $20^{th}$ is corrected as the model incorporates newer data. \ref{['et_p_pred']}: Illustrates the estimated probability transition rates over the same period. The black line shows the median probability transition, while the shaded area represents the $97.5^{th}$ percentile, indicating the uncertainty in these predictions. The probability transition rates exhibit increased uncertainty between the second virus mutation on November $15^{th}$ and vaccination on December $20^{th}$. This uncertainty is reflected in changing contributing factors, with the median estimate showing a fluctuating trend.
• Figure 2: Proposed model estimated values for $\omega_t$ for 9/1/20 - 1/21/21 period. Before the October $25^{th}$, 2020 lockdown, the average contact rate rose following a mutation, but after the lockdown, it dropped and remained steady until a second lockdown around Christmas.
• Figure 3: \ref{['fig:I_t_2020-09-01_2020-10-01']}-\ref{['fig:I_t_2020-10-20_2020-11-19']} display the estimated values of $I_t$ using Bayesian hierarchical changing point couple with 31 days sliding windows from 9/1/2020 - 11/19/2020 period. The green dots represent the observed number of infected individuals, while the blue markers show the model's estimates of infections. The white region indicates the 31-day window used to estimate model parameters, and the yellow area represents the week-long forecast based on these estimates. Vertical red lines mark the date of the mutation considered in the model, the green line shows when vaccination started, and the blue line represents the lockdown date. The sliding window approach iteratively adjusts predictions by incorporating new data every 7 days, generating median estimates to forecast future infection numbers. The results capture the overall trend in predicting the number of infections, with the median estimates closely matching the observed data and accurately reflecting fluctuations associated with key events during this period.
• Figure 4: \ref{['fig:I_t_2020-10-27_2020-11-26']}-\ref{['fig:I_t_2020-12-15_2021-01-14']} display the estimated values of $I_t$ using Bayesian hierarchical changing point couple with 31 days sliding windows from 10/27/2020 - 1/21/2021 period. The green dots represent the observed number of infected individuals, while the blue markers show the model's estimates of infections. The white region indicates the 31-day window used to estimate model parameters, and the yellow area represents the week-long forecast based on these estimates. Vertical red lines mark the date of the mutation considered in the model, the green line shows when vaccination started, and the blue line represents the lockdown date. The sliding window approach iteratively adjusts predictions by incorporating new data every 7 days, generating median estimates to forecast future infection numbers. The results capture the overall trend in predicting the number of infections, with the median estimates closely matching the observed data and accurately reflecting fluctuations associated with key events during this period.
• Figure 5: Directional error estimation ($DNE$) of $I_t$ for 9/1/20 - 1/21/2020. Around mutation dates, $DNE$ values increase, indicating that the model overestimates infection numbers due to challenges in adapting to new viral strains. During lockdown periods, $DNE$ decreases, often showing underestimation, while after the vaccination roll-out on December, $DNE$ stabilizes, suggesting better model alignment with observed data.