RAPID: Active Tracking of Disease Spread in CoVID19 via Graph Predictive Analytics

冠状病毒病2019(新冠肺炎)已成为一场全球性的公共卫生危机。截至2020年4月10日，180多个国家和地区约有170多万人确诊为新冠肺炎病例，死亡人数超过10万人。在美国，确诊病例超过50万例，死亡人数近2万人，而且这些数字还在继续大幅上升。随着疫情的发展，显然迫切需要确保基础设施和关键服务的可获得性。目前控制疫情的计划是基于完善的流行病预测“隔间”模型的预测。这些模型依赖于基于同质种群、同质混合和几个关键超参数(如基本繁殖率)的假设的微分方程式。传染病和流行病管理专家都知道，将观测数据与此类模型的参数相匹配，是一种描述流行病学特征的做法，而不是生成有效和可操作的预测。因此，迫切需要大幅更新这些模型，以考虑到从多个数据来源和地点收集的实地数据。这在工程先发制人的干预措施以遏制疾病传播方面尤其相关。目前的COVID疾病数据是以地理空间格式组织的，即按地理位置索引的感染、死亡和疑似病例，范围从市、县或州一级的粗略程度。该项目旨在开发和演示一些技术，这些技术利用数据的地理空间性质、疾病统计数据(以及预测)的时间演变和多源数据的综合，以帮助快速和先发制人地将可用的医疗资源分配到最需要的地区。对新冠肺炎疫情进行建模和设计干预措施是重大挑战。这个项目通过图表分析的镜头来看待这个问题。特别是，它寻求利用感兴趣的地理空间区域之间的相似性信息来改进流行病预测，并帮助设计有效的干预措施。作为第一步，流行病预测问题被建模为从低维观测重建高维动力系统。通过利用感兴趣的地区之间的相似性信息，将加强对这样一个模型的估计。虽然地理空间接近度图是相似性图的自然候选者，但它未能基于其他因素，如人口的社会学和生物特征，捕捉地理区域之间的长期统计依赖关系。使用图形建模的技术，该项目将开发新的技术，以便在持续的大流行期间为流行病建模学习具有统计意义的图表。此外，生成的准确的时间序列预测将与基于图的相似性度量相结合，以设计有效的干预措施来遏制疫情的传播。目前正在使用随机公式和新出现的方法在带有时间序列观测的图表上进行异常检测；基于这些范例的最佳政策将转化为针对不断演变的大流行的干预战略。该项目利用与马里科帕县和亚利桑那州当地社区利益攸关方的伙伴关系，通过弹性知识交换(KER)来实施所开发的方法，并确保其技术进步能够产生可在全国和全球推广的有意义的见解。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Corona Virus Disease 2019 (COVID-19) has emerged as a public health crisis of global proportions. As of April 10, 2020, there are approximately 1.7 million confirmed COVID-19 cases in more than 180 countries, with over 100,000 deaths. In the US, there are more than 500,000 confirmed cases and nearly 20,000 fatalities, and these numbers are continuing to rise sharply. There is a clear and acute need for ensuring the availability of infrastructure and critical services as the epidemic progresses. Current plans for controlling the epidemic are based on forecasts from well established “compartment” models for epidemic prediction. These models rely on differential equations based on assumptions of homogeneous populations, homogeneous mixing, and knowledge of several critical hyperparameters such as the base reproduction rate. It is well known among experts in infectious diseases and epidemic management that fitting observed data to the parameters of such models is an exercise in characterizing the epidemiology as opposed to generating valid and actionable predictions. Consequently, there is an urgent need to significantly update these models to account for the data collected on the ground from multiple data sources and locations. This is especially relevant in engineering preemptive interventions to check disease spread. Current COVID disease data are organized in a geospatial format, i.e., infected, deceased, and suspected cases indexed by geolocation, which can range from city-, county-, or state-level coarseness. This project aims to develop and demonstrate techniques that use the geospatial nature of the data, the temporal evolution of disease statistics (along with predictions), and synthesis of multiple sources of data to help rapidly and preemptively allocate available medical resources toward the areas of greatest need.  Modeling the COVID-19 epidemic and designing interventions are significant challenges. This project looks at the problem through the lens of graph analytics. In particular, it seeks to use similarity information between geospatial regions of interest to improve epidemic predictions and to design effective interventions. As a first step, the problem of epidemic prediction is being modeled as the reconstruction of a high-dimensional dynamical system from low-dimensional observations. The estimates of a model thus learned will be enhanced by leveraging similarity information between the localities of interest. While the geospatial proximity graph is a natural candidate for the graph of similarities, it fails to capture long-range statistical dependencies between geographical regions based on other factors such as the sociological and biological features of a population. Using techniques from graphical modeling, this project will develop new techniques for learning statistically meaningful graphs for epidemic modeling during an ongoing pandemic. Furthermore, the accurate time-series prediction generated will be combined with the graph-based similarity measures to design effective interventions to check the spread of the epidemic. This is being approached using a stochastic formulation and emerging methods for anomaly detection on graphs with time series observations; optimal policies based on these paradigms will be translated into interventional strategies for an evolving pandemic. The project leverages partnerships with local community stakeholders in Maricopa County and the State of Arizona through the Knowledge Exchange for Resilience (KER) to implement the methodologies developed, and to ensure its technical advances can produce meaningful insights that can generalize nationally and globally. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.