Smart and scalable epidemic prediction and control

新型传染病，如非典和新冠肺炎，是对公众健康的主要威胁。在缺乏疫苗或先前免疫的情况下，以及在这些疾病出现时围绕其特征存在很大的不确定性，模型构成了我们的第一道防线。数学模型是一种计算工具，它将我们关于疾病如何传播的知识与现有的流行病数据结合起来，例如每天的病例计数。模型可以提供对推动传播的关键因素的了解，对即将到来的病例或死亡的预测，以及对非药物干预(NPI)潜在影响的估计，例如社会距离或封锁。鉴于这种预测能力，模型产出经常被用作政策制定的证据。然而，这一证据的可靠性可能在很大程度上甚至有时出人意料地取决于模型的规模或详细程度。流行病是复杂的现象，涉及在地理、人口和其他尺度上传播的差异，称为异质性。模拟所有这些异质性的精细模型可能会产生不可靠的预测，因为我们可能只有关于每个差异的稀缺数据，并且必须做出更多假设才能使用该模型。粗略的模型对整个国家的这些异质性进行平均，或者由于基于年龄的风险而忽略差异，这些模型可能更容易使用，但过于自信，只能证明停摆等NPI是迟钝的。选择合适的规模来对传染病进行建模和应对是流行病学的前沿问题。错误的规模可能会误导政策，使应对大流行的风险、成本和效率低下。两个主要问题使这一模型选择问题从根本上具有挑战性。首先，最可靠的建模尺度(例如，地方、区域或国家)因地点、时间和对NPI的反应而不同。现有的模型很少适应这种波动，或者当它们确实适应时，往往会变得非常复杂。其次，嘈杂的数据、政策的后勤限制(例如，学校停课可能只发生在全区范围内)以及人们的反应行为对模型施加了很大程度上未知的性能限制，限制了预测的范围或NPI的效率。我将开发智能模型来解决这些问题。通过将简单的流行病模型连接到层级或组中，其中每个较低级别的组描述了一些感兴趣的异质性，而每个较高级别的组描述了这种异质性的平均值，我的目标是构建新的模型，真实地描述流行病的许多相互作用的规模。信息论和分散控制理论是工程领域，它们提供了减少不确定性和管理反应环的独特而严格的方法，而这些方法在流行病学中很少使用。通过将这些领域的原则与政策科学家的专家意见结合起来，我将设计新的算法，重新构建这些层次结构，以揭露并绕过性能限制，并在任何时候准确地确定最可靠的规模，以实际抗击流行病。这些智能框架智能地平衡了传播的细节和现有数据，以可靠地了解这些细节，将推动流行病建模的界限。将它们应用于SARS和新冠肺炎的不同数据集，我将(I)得出强有力的传播早期预警指标(例如，预测流行病是否可能出现第二波疫情的迹象)，(Ii)提高对建模限制如何转化为对预测或控制疫情爆发能力的限制的理解，(Iii)制定新的战略，以协调不同规模的NPI，以提高未来应对大流行的效率(例如，发现何时地方封锁的组合可能比全国封锁更有效)。适应流行病不断变化的现实的智能模型可以巩固可靠和更知情的公共卫生政策的证据基础。

Novel infectious diseases, such as SARS and COVID-19, are pre-eminent threats to public health. In the absence of vaccines or prior immunity and the presence of large uncertainties surrounding the characteristics of these diseases when they emerge, models form our first line of defence. Mathematical models are computational tools that combine our knowledge of how diseases spread with available epidemic data, for example daily counts of cases. Models can provide understanding of the key factors driving transmission, forecasts of upcoming cases or deaths and estimates of the potential impact of non-pharmaceutical interventions (NPIs), such as social distancing or lockdowns. Given this predictive power, model outputs often serve as evidence for policymaking. However, the reliability of this evidence can depend substantially and sometimes unexpectedly on the scale or level of detail of the model. Epidemics are complex phenomena, involving differences, called heterogeneities, in spread across geographic, demographic and other scales. Fine scale models simulating all of these heterogeneities may yield unreliable forecasts because we may only have scarce data on each difference and have to make more assumptions to use the model. Coarse models, which average these heterogeneities over an entire country or ignore differences due to age-based risks, may be easier to use but overconfident and only able to evidence blunt NPIs such as lockdowns. Selecting the right scale at which to model and respond to infectious diseases is a problem at the forefront of epidemiology. Getting this scale wrong could misinform policy, making pandemic response risky, costly and ineffective.Two main issues make this model selection problem fundamentally challenging. First, the most reliable scale for modelling (e.g. locally, regionally or nationally) varies with location, time and response to NPIs. Existing models rarely adapt to this fluctuation or when they do tend to become very complex. Second, noisy data, logistical constraints on policy (e.g. school closures may only occur district-wide) and the reactive behaviours of people impose largely unknown performance limits on models, restricting the horizons of forecasts or efficiency of NPIs. I will develop smart models to resolve these issues. By connecting simple epidemic models into hierarchies or groups, where each lower-level group depicts some heterogeneity of interest and each higher one averages over that heterogeneity, I aim to construct novel models that realistically describe the many interacting scales of pandemics. Information theory and decentralised control theory are engineering fields that offer unique and rigorous ways of mitigating uncertainty and managing reactive loops that are seldom used in epidemiology. By combining principles from these fields together with expert input from policy scientists, I will design new algorithms that restructure these hierarchies to expose and bypass performance limits, and to pinpoint the most reliable scales for practically combatting pandemics at any time. These smart frameworks, which intelligently balance the details of spread with the available data to reliably learn about those details, will push the boundaries of epidemic modelling. Applying them to diverse SARS and COVID-19 datasets, I will (i) derive robust early-warning indicators of transmission (e.g. signs that foretell if an epidemic might have a second wave), (ii) improve understanding of how limits to modelling translate into restrictions on how well we can predict or control outbreaks and (iii) derive new strategies for coordinating NPIs across different scales to improve the efficiency of future pandemic response (e.g. discovering when combinations of local lockdowns might be more effective than a national one). Smart models, which adapt to the changing reality of pandemics, can solidify the evidence base for reliable and better-informed public health policy.