Merging machine learning and mechanistic models to improve prediction and inference in emerging epidemics

项目摘要 当一种既有或新出现的传染病爆发时，我们会问一系列标准问题， 对拯救生命的公共卫生反应至关重要：未来的发病率将发生在哪里？有多少案件 有吗？我们在哪里可以最有效地干预？本研究的动机是基于真实的世界 回答这些问题对于做出实际的公共卫生决策至关重要， 从决定2014-15年是否以及在哪里建立更多的埃博拉治疗单位， 西非埃博拉疫情，以确定应在西非地区使用口服霍乱疫苗的优先地区， 2016-17年也门霍乱疫情，选择可能发生足够病例的地点， 优先采取干预措施，减缓COVID-19在全球的传播。这些决定所依据的预测是 通常使用依赖于疾病传播知识的流行病模型生成 机制和流行病理论或使用统计模型来预测基于 协变量与观察计数之间的关系。然而，这两种方法都受到限制， 特别是在流行病的早期，那时观察到的病例很少。这个项目是基于总体 科学前提是结合了机械流行病模型和统计学的优势的推论，联合收割机 协变量模型在预测和决策方面将大大优于单独的任何一种方法， 应对新出现的传染病威胁。具体而言，本项目旨在（1）制定一个框架， 预测正在发生的疾病爆发的发病率，融合了机械和机器学习方法; (2)使用回顾性数据验证框架，并应用该框架为决策提供信息 在新出现的流行病中;（3）将这种推理预测框架融入因果决策理论 优化公共卫生应对新出现的流行病的关键行动;以及（4）制定 用于传染病流行预测和决策分析的可访问和可扩展工具。 我们将通过严格的模拟研究和应用所提出的方法来验证这些方法 到最近重要流行病的回顾性数据（例如，埃博拉，霍乱和COVID-19，如上所述 以上）。我们将前瞻性地应用我们的方法来为应对新出现的疾病威胁提供信息， 在项目期间发生的任何事件，包括正在进行的COVID-19大流行。确保开发的工具 是有用的，高效的，用户友好的，我们将与国际人道主义组织合作， 流行病成功地完成这些目标将为预测提供一个灵活和有效的框架 在流行病持续期间进行决策，同时允许在机械和统计方面进行创新 接近。在这样做的过程中，它将提供各种工具，以优化应对措施，并减少发病率和死亡率， 公共卫生危机。

PROJECT SUMMARY When an outbreak of an established or emerging infectious disease occurs we ask a standard set of questions that are critical to a lifesaving public health response: Where will future incidence occur? How many cases will there be? And where can we most effectively intervene? The proposed research is motivated by real world instances where answering these questions was critical to making practical public health decisions, and current methods came up short: from deciding if and where to build additional Ebola Treatment Units in the 2014-15 West African Ebola epidemic, to identifying priority districts where oral cholera vaccine should be used in the 2016-17 cholera outbreak in Yemen, to picking locations where sufficient cases might occur to selecting and prioritizing interventions to slow the spread of COVID-19 worldwide. Forecasts informing such decisions are typically generated either using an epidemic model that relies on knowledge of the disease transmission mechanism and epidemic theory or using a statistical model to project the expected number of cases based on the relationship between covariates and observed counts. However, both approaches are subject to limitations, particularly early in an epidemic when few cases are observed. This project is based on the overarching scientific premise that inferences that combine the strengths of mechanistic epidemic models and statistical covariate models will substantially outperform either approach alone in forecasting and making decisions to confront emerging infectious disease threats. Specifically, this project aims to (1) Develop a framework to forecast incidence in ongoing outbreaks that merges mechanistic and machine learning approaches; (2) Validate the framework using retrospective data and apply the framework to inform decision making in emerging epidemics; (3) Integrate this inferential forecasting framework into causal decision theory to optimize critical actions in the public health response to emerging epidemics; and (4) Develop accessible and extensible tools for forecasting and decision analysis in infectious disease epidemics. We will validate these approaches using rigorous simulation studies and by applying the proposed approaches to retrospective data from important recent epidemics (e.g., Ebola, Cholera and COVID-19, as mentioned above). We will prospectively apply our approach to inform the response to emerging disease threats that occur during the project period, including the ongoing COVID-19 pandemic. To ensure that the tools developed are useful, efficient, and user friendly, we will work with international humanitarian organizations responding to epidemics. Successful completion of these aims will provide a flexible and validated framework for forecasting and decision making during ongoing epidemics, while allowing for innovation in mechanistic and statistical approaches. In doing so it will provide tools to optimize responses and reduce morbidity and mortality during public health crises.