A mathematically-driven framework for pandemic planning and management

尽管新冠是过去20年来的第四次全球大流行(2012-13年的MERS、2009-10年的H1N1、2002-04年的SARS)，但新冠肺炎让人们注意到缺乏定义明确和强有力的工具来管理大范围的流行病。这项研究计划通过设计最佳缓解策略和疫苗优先政策，满足加拿大公共卫生机构和医疗保健机构在管理COVID和未来大流行方面的需求。为了获得这些最优策略，这项研究利用我们完整的基于COVID代理的模拟模型(ABM)，在适用于任何人群和疾病的框架内进行深入的、数学驱动的分析。公共卫生机构主要依靠高水平的疾病传播预测模型，称为分隔模型(通常是传统的易感-感染-康复(SIR)模型的扩展)，以估计大流行将如何根据每天的病例、住院和死亡数量传播。这些模型的好处是，它们只需要有关疾病的高级信息，例如繁殖数(R0)(单个病例造成的新感染的平均数量)和人口，例如大小和接触率；对最小数据的需求意味着这些模型可以在大流行的早期阶段迅速发展。缺点是，只有高级别的信息才会进入，所以只有高级别的信息会出来，这使得这些模型不适合进行详细的政策评估。通过包括详细的人口统计、医疗和经济信息，ABM可以更准确地模拟疾病传播，测试和优化细微差别的缓解策略。特别是，我们根据地区流行情况纳入个人健康状况(例如，共病)，允许对人口结果进行新的调查，并公平地考虑到人口多样性。例如，农村和较贫穷的社会经济地区的健康状况通常较差，合并症的患病率较高，获得医疗保健的机会也较少；这些地区将受到流行病的更大打击，不会出现在区隔模型中。应对新出现的大流行的缓解战略是由公共卫生政策制定者根据经验、直觉、有限的数据分析(由于大流行的新颖性)以及咨询建模专家来回答有关潜在缓解战略的“假设”问题而制定的，而这些战略几乎肯定不会包含最佳战略。这项研究超越了这种被动的方法，通过ABM及其联系网络上的优化和机器学习方法的组合，直接优化策略，而不是以特别的假设方式。此外，这项工作提供了对公共卫生官员可以拉动的政策“杠杆”的有效性的人类可解释的评估。

COVID-19 has brought into focus the lack of well-defined and robust tools to manage widespread pandemics, despite the fact that COVID is the fourth major global pandemic in the past 20 years (MERS in 2012-13, H1N1 in 2009-10, SARS in 2002-04). This research program addresses the needs of Canada's public health agencies and healthcare institutions in managing COVID and future pandemics by designing optimal mitigation strategies and vaccine prioritization policies. To obtain these optimal policies, this research leverages our completed COVID agent-based simulation model (ABM) for deep, mathematically-driven analysis in a framework generalizable to any population and disease. Public health agencies predominantly rely on high-level disease spread prediction models, called compartmental models (typically extensions of the traditional susceptible-infectious-recovered (SIR) model), to estimate how the pandemic will spread in terms of daily number of cases, hospitalizations, and deaths. The benefit of these models is that they only require high-level information about the disease, e.g., reproduction number (R0) (the average number of new infections caused by a single case), and population, e.g., size and contact rate; the need for minimal data means these models can be rapidly developed in the early stages of a pandemic. The drawback is that only high-level information goes in, so only high-level information comes out, making these models ill-suited for detailed policy assessments. By including detailed population demographic, medical, and economic information, ABMs, where individuals and their unique characteristics are individually represented, can more precisely simulate disease spread and test and optimize nuanced mitigation strategies. In particular, our incorporation of individual health status (e.g., comorbidities) based on regional prevalence allows for novel investigation of population outcomes, as well as equitably accounts for population diversity. For example, rural and poorer socio-economic areas usually have worse health status and higher prevalence of comorbidities, as well as reduced access to healthcare; these areas will be hit harder by pandemics, and are not represented in compartmental models. Mitigation strategies to respond to an emerging pandemic are generated by public health policymakers based on experience, intuition, limited data analysis (due to novelty of the pandemic), and consultation with modelling experts to answer "what if" questions regarding potential mitigation strategies, which almost certainly will not contain the optimal strategy. This research goes beyond this reactive approach by optimizing policies directly, rather than in an ad hoc what-if fashion, through a combination of optimization and machine learning approaches on the ABM and its contact networks. Additionally, this work provides human-interpretable assessments of the effectiveness of policy "levers" that can be pulled by public health officials.