Evaluating COVID-19 vaccine allocation policies using Bayesian $m$-top exploration

TL;DR

The paper tackles the problem of efficiently evaluating fine-grained COVID-19 vaccine allocation policies under substantial computational cost and decision uncertainty. It introduces a Bayesian anytime $m$-top exploration framework, implemented as epidemic bandits, to identify the top $m$ vaccination strategies with quantified uncertainty. Through ground-truth validation in a Belgian STRIDE-based setting, the authors show how different contact-reduction regimes and vaccine-uptake levels shape the prioritization of age groups and vaccine types, and they provide a GPL-licensed software framework for replication and future use. The work offers policymakers a compact set of high-utility strategies with uncertainty bounds, enabling robust, data-informed decisions during evolving epidemics.

Abstract

Individual-based epidemiological models support the study of fine-grained preventive measures, such as tailored vaccine allocation policies, in silico. As individual-based models are computationally intensive, it is pivotal to identify optimal strategies within a reasonable computational budget. Moreover, due to the high societal impact associated with the implementation of preventive strategies, uncertainty regarding decisions should be communicated to policy makers, which is naturally embedded in a Bayesian approach. We present a novel technique for evaluating vaccine allocation strategies using a multi-armed bandit framework in combination with a Bayesian anytime $m$-top exploration algorithm. $m$-top exploration allows the algorithm to learn $m$ policies for which it expects the highest utility, enabling experts to inspect this small set of alternative strategies, along with their quantified uncertainty. The anytime component provides policy advisors with flexibility regarding the computation time and the desired confidence, which is important as it is difficult to make this trade-off beforehand. We consider the Belgian COVID-19 epidemic using the individual-based model STRIDE, where we learn a set of vaccination policies that minimize the number of infections and hospitalisations. Through experiments we show that our method can efficiently identify the $m$-top policies, which is validated in a scenario where the ground truth is available. Finally, we explore how vaccination policies can best be organised under different contact reduction schemes and we investigate the impact of vaccine uptake proportions (i.e., the proportion of individuals that will comply with the strategy and take the vaccine).

Figures (23)

• Figure 1: Posteriors for an artificial bandit ($K = 6$, $m = 3$) (gray) and BFTS’ decision boundary (red) with confidence bounds to demonstrate its uncertainty
• Figure 2: (a) Stacked bar chart of the reported vaccine supply from January 1st 2021 Vaesen2022. (b) Stacked bar chart for one example of an uptake strategy where the entire populations accepts vaccines, starting with vaccinating children and young adults with vector-based vaccines and youngsters with mRNA vaccines. When the youngsters are fully vaccinated, remaining and newly arrived mRNA vaccines will be allocated first to other groups prioritized for mRNA vaccines, if any (none in this example). Subsequently, vaccines will be distributed to other age groups without prioritization, specifically adults and the elderly in this example.
• Figure 3: Ground truth of all arms for the baseline scenario, ranked based on 100 stochastic simulations. Ground truth of all arms for the baseline scenario, ranked based on infections (1 - ARI) and hospitalisations (1 - ARH). We note that the ranking based on infections versus hospitalisations does not necessarily match, here we show an independent ranking for each of the criteria.
• Figure 4: Ground truth of the top-$10$ vaccination strategies for the baseline scenario when minimising the infection (ARI) and hospitalisation (ARH) attack rates. Each strategy in the top-$10$ strategies is represented by 5 numbered circles, each representing a specific age group as highlighted in the legend. The colour of the circle indicates which vaccine type is being prioritised for the given strategy. For example, the first strategy when optimising for ARI prioritises vector-based vaccines for children and young adults. Youngsters receive priority for an mRNA vaccine, while adults and elderly receive no vaccine priority.
• Figure 5: Learning curves for the ground truth based on infections (ARI) and hospitalisations (ARH), top vs bottom row, respectively. Left column: The average proportions of correctly ranked arms, with standard deviation. Right column: The average sum of true means, with standard deviation.

Theorems & Definitions (2)

• Definition 1: Multi-armed bandit
• Definition 2: Stochastic epidemiological model