IIII: RAPID: Interventional COVID-19 Response Forecasting in Local Communities Using Neural Domain Adaptation Models

关于新冠肺炎的传播，以及我们的缓解策略是如何影响传播的，我们仍然有很多不了解的地方。人口结构、人口密度、商业结构和社会文化因地区而异。将这些当地因素与感染人数和医院资源的可用性联系起来，可以为当地政策制定者提供宝贵的科学和数据驱动的指导。与现有的经典流行病模型不同，在这个项目中，我们的目标是基于尖端人工智能技术建立新的预测模型。其目标是提供管理员所需的及时、本地化的信息，以便战略性地分配资源和规划重新开业。我们方法的一个关键优势是它能够将来自新冠肺炎病例较多地区的数据与每个当地社区特有的美国人口普查微观数据结合起来，从而帮助我们对政策决策的局部影响做出细粒度预测。现有的新冠肺炎病例预测模拟模型要么忽略了当地社区细粒度的人口、社会和文化差异，要么往往需要复杂的手动参数设置来估计干预效果。另一方面，现有的统计模型需要有大量的数据可用，因此不能在每个地方一级获得足够可信的预测。我们提出了一种完全不同的方法，它建立在最新的神经网络模型(如变形金刚)上，以克服这些弱点。该方法进行领域自适应和少镜头学习，使得从其他区域学习到的知识即使在只有几个数据点的情况下也可以适应于本地社区。具体地说，我们的方法将在圣巴巴拉平房医院的合作者的临床指导下，创造性地从美国人口普查的美国社区调查数据、国内外其他地区的新冠肺炎相关数据以及其他相关类型的流行病中提取信息。这一奖项反映了美国国家科学基金会的法定使命，并已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。

There is still much we do not understand about the spread of COVID-19, and how our mitigation strategies are affecting the spread. Demography, population density, business structure, and social culture differ across regions. Correlating these local factors with the number of infections and the availability of hospital resources can provide precious scientific and data-driven guidance to local policy makers. Different from existing, classic epidemic models, in this project we aim to build novel forecasting models based on cutting-edge AI techniques. The goal is to provide timely, localized information needed by administrators for strategic allocation of resources and planning towards reopening business. One key advantage of our approach is that it is able to combine the data from regions with more COVID-19 cases with the US Census microdata that characterize each local community, hence helping us to make fine-grained predictions of the localized effects of a policy decision.Existing simulation models for COVID-19 cases forecasting either ignore the fine-grained demographical, social and cultural difference at local communities, or often require complicated, manual parameter setting for estimating the effect of interventions. Existing statistical models, on the other hand, require substantial amount of data to be available, hence are not able to obtain sufficiently confident predictions on each local level. We propose a fundamentally different approach that is built on the newest neural network models like Transformers to overcome these weaknesses. The proposed approach performs domain adaption and few shot learning, so that knowledge learned from other regions can be adapted to local communities even when only a few data points are available. Specifically, our approach will creatively draw information from the US Census’s American Community Survey data, COVID-19 related data from other regions at home and abroad, as well as other related kinds of epidemics under the clinical guidance of our collaborators from the Santa Barbara Cottage Hospital.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.