CAREER: Big Computation and the Management of Emerging Infectious Diseases

新发传染病占全球疾病负担的25%以上，占全球死亡人数的32%以上。目前的eid，如中东呼吸综合征冠状病毒（MERS）和耐抗生素超级细菌，有可能对公众健康造成毁灭性影响。本项目正在开发的方法可用于将关于艾滋病的实时数据转化为关于在何处、何时以及向谁实施干预措施的建议，以便尽量减少该疾病的负面影响，同时减少总体资源消耗。此外，这些建议被设计成在主题上下文中可立即解释，从而使决策者能够将来自复杂和异构数据流的信息纳入疾病管理。应用这些方法有可能以比现有管理计划更低的成本降低死亡率和发病率。此外，支持干预建议的模型将产生关于EID动态的新知识。该研究项目旨在为在线顺序决策做出基础性贡献，并为数据驱动的eid管理创建一个新的统计框架。我们将EID概念化为在有限的一组位置上传播，这些位置可能是空间中的物理位置或网络中的节点。分配策略将EID的管理形式化，并由一系列功能（每个干预决策一个）表示，这些功能将EID的最新信息映射到建议治疗的位置子集。最优分配策略在EID期间使某些平均效用函数最大化。从EID上的数据构建最优分配策略具有挑战性，因为：(i)分配的数量在位置数量上呈指数级增长；（ii）评估和管理必须同时进行；（iii）空间邻近导致因果干扰；（iv）分配策略必须能够为主题专家所理解。我们整合了统计学、计算机科学、优化和疾病生态学的思想来克服这些挑战。将仿真优化算法与策略搜索算法相结合，构造了最优分配策略的在线估计器；这一策略权衡了探索分配选择，以改善疾病动态的估计，并利用当前估计的动态来立即减缓EID的传播。我们证明了治疗分配问题可以被转化为一个无限维的强盗问题。我们利用这种联系来推导估计算法，这些算法可以扩展到非常大的分配问题，并且可以进行理论研究。我们将策略搜索和基于强盗的估计器与一类新的分配策略结合起来，这些分配策略可以表示为一系列if-then语句，这些语句可以立即被主题专家解释，并且可以根据专家判断随时进行调整。在该类中，我们导出了估计分配策略的近似误差的非参数下界；该界用于对估计的最优分配策略执行拟合优度检验。

Emerging infectious diseases (EIDs) account for more than 25% of global disease burden and more than 32% of global deaths. Current EIDs like Middle East Respiratory Syndrome Coronavirus (MERS) and antibiotic-resistant superbugs have the potential to make devastating impacts on public health. The methodologies under development in this project can be used to translate real-time data on EIDs into recommendations about where, when, and to whom to apply interventions so as to minimize negative impacts of the disease while reducing overall resource consumption. Furthermore, these recommendations are designed to be immediately interpretable in a subject matter context, thereby empowering decision makers to incorporate information from complex and heterogeneous data streams into disease management. Application of these methodologies has the potential to reduce mortality and morbidity at lower cost than existing management plans. Furthermore, models underpinning intervention recommendations will generate new knowledge about EID dynamics. This research project aims to make fundamental contributions to online sequential decision making and to create a new statistical framework for data-driven management of EIDs. We conceptualize the EID as spreading across a finite set of locations, which might be physical locations in space or nodes in a network. An allocation strategy formalizes management of an EID and is represented by a sequence of functions, one per intervention decision, that map up-to-date information on an EID to a subset of locations recommended for treatment. An optimal allocation strategy maximizes some mean utility function over the duration of the EID. Construction of an optimal allocation strategy from data on an EID is challenging because: (i) the number of allocations is exponential in the number of locations; (ii) estimation and management must occur simultaneously; (iii) spatial proximity induces causal interference; and (iv) an allocation strategy must be interpretable to subject matter experts. We integrate ideas from statistics, computer science, optimization, and disease ecology to overcome these challenges. We combine simulation-optimization with policy-search algorithms to construct an online estimator of the optimal allocation strategy; this strategy trades off exploring allocation choices that improve estimates of disease dynamics with exploiting current estimated dynamics to immediately slow spread of the EID. We show that the treatment allocation problem can be recast as an infinite-dimensional bandit problem. We leverage this connection to derive estimation algorithms that scale to very large allocation problems and are amenable to theoretical study. We combine our policy-search and bandit-based estimators with a novel class of allocation strategies that can be expressed as a sequence of if-then statements that are immediately interpretable to subject-matter experts and can be readily adjusted based on expert judgment. We derive a non-parametric lower bound on the approximation error of an estimated allocation strategy within this class; this bound is used to perform goodness-of-fit tests for the estimated optimal allocation strategy.