Spatiotemporal forecasting of COVID-19 by integrating machine learning and epidemiological modeling

项目总结/摘要 在这场持续的COVID-19大流行中，准确和早期预测疾病的传播至关重要。 传染性很强的SARS-CoV-2对大流行形势和未来趋势的正确预测， 有效的资源分配和政府政策，以减少COVID-19对公众的不利影响 健康和经济。虽然各种流行病学模型有助于预测感染 传播，他们的整体平均方法在很大程度上忽略了异质性中的关键信息。 传统的流行病学模型主要关注全球平均趋势，分析能力有限 由于大流行的空间异质性，无法进行局部预测 situations.鉴于COVID-19大流行的快速变化，采取紧急应对措施也具有挑战性 新的流行病学数据，如果我们仅仅依靠人类的智慧。最近，机器学习（ML） 取得了巨大的进步，并表明计算机在分析复杂的 高维数据集。我们的实验室一直在通过开发ML来解决细胞生物学中的这些挑战。 在亚细胞水平上进行荧光活细胞图像分析的平台。我们建立了一种方法， 对细胞突起时间序列的亚细胞异质性进行反卷积， 具有不同药物敏感性的突出表型。因此，我们的目标是利用我们的ML平台， 解决流行病学建模中的这些技术挑战，以快速预测COVID-19的传播， 在美国的县一级。首先，我们将推进我们的ML平台，用于空间数据的反卷积。 COVID-19动态的异质性。该方法将流行病学模型和ML相结合， 美国各县的集群共享类似的时间模式。其次，我们将应用我们基于深度学习的 特征学习，其中深度神经网络学习由先验知识指导的关键特征， 流行病学的数学模型这将使我们能够生成细粒度的预测 COVID-19传播的地图。我们的ML平台将为流行病学带来前所未有的预测能力， 使我们能够对当前的COVID-19大流行和未来的其他传染病采取紧急应对措施 爆发。

PROJECT SUMMARY/ABSTRACT In this ongoing COVID-19 pandemic, it is crucial to have an accurate and early prediction of the spread of highly infectious SARS-CoV-2. A correct prediction of the pandemic situation and future trends enables effective resource allocation and government policies to reduce the detrimental effect of COVID-19 on public health and economics. Although various epidemiological models have facilitated the prediction of the infection spread, their ensemble-averaging approach largely disregards critical information within the heterogeneity. Conventional epidemiological modeling is focused on the global average trends, which is limited in analyzing dynamic local information and does not allow local prediction due to spatial heterogeneity of the pandemic situations. Given the rapid changes of the COVID-19 pandemic, it is also challenging to take urgent responses to the new epidemiological data if we solely rely on human intelligence. Recently, machine learning (ML) is making tremendous progress and has shown that computers can outperform humans in analyzing complex high-dimensional datasets. Our lab has been addressing these challenges in cell biology by developing an ML platform for fluorescence live cell image analyses at the subcellular level. We established the method to deconvolve the subcellular heterogeneity of time series of cell protrusion, which identified distinct subcellular protrusion phenotypes with differential drug susceptibility. Thus, our goal is to leverage our ML platform to address these technical challenges in epidemiological modeling for rapid forecasting of COVID-19 spread at the county level in the United States. First, we will advance our ML platform for the deconvolution of spatial heterogeneity of COVID-19 dynamics. This method will integrate epidemiological models and ML to identify the clusters of US counties sharing similar temporal patterns. Second, we will apply our deep learning-based feature learning, where the deep neural networks learn the critical features guided by prior knowledge and well-established epidemiological mathematical models. This will allow us to generate fine-grained forecasting maps of Covid-19 spread. Our ML platform will bring unprecedented prediction power to epidemiology and enable us to take urgent responses to the current COVID-19 pandemic and future other infectious disease outbreaks.