Integrating Air Pollution Prediction Models: Uncertainty Quantification and Propagation in Health Studies

项目摘要 环境空气污染是全球环境威胁，导致数百万人死亡，数百人死亡 每年数百万的伤残调整寿命年(DALY)。然而，空气污染健康的主要限制 研究仍然是暴露评估。尽管在空气污染评估方面取得了很大的进展 已经开发了几个复杂的时空模型来预测每天的空气污染水平 研究对象的住址，这些模型的表现在空间和时间上有所不同。 然而，即使是平均性能最好的预测模型，在一定程度上也会具有有限的预测能力 空间和时间点。此外，与使用单一预测模型相关的不确定性一直是 在健康研究中一直被忽视，这可能导致对健康影响估计的无效推断，以及 不同的研究发现不一致。我们建议通过开发一种新的乐团来解决这一关键差距 空气污染健康研究中暴露评估的模型框架，整合信息 多种现有的预测模型。有了这种方法，我们将首次能够全面地 量化与环境空气污染暴露相关的任何模型间和模型内的不确定性。我们将发展 单污染物和多污染物环境下的综合方法。我们建议将开发的方法应用于 并利用两个全国性的公开队列充分传播健康影响评估中的暴露不确定性，主要是 医疗保险和医疗补助，以及纽约州(全州范围)住院人数的公开队列 规划与研究合作系统(SPARCS)。这些数据集提供了有关以下内容的信息 全美所有老年、低收入和残疾美国人(联邦医疗保险和医疗补助， 分别)，居住信息邮政编码水平，以及纽约州98%的住院治疗， 在住址上有相关信息。具体来说，我们将评估长期和短期影响 空气污染暴露对死亡率(医疗保险和医疗补助)和心肺疾病发病率(全部三项)的影响 我们交流空气污染预测，空气污染暴露的时空不确定性 向公众和监管机构提供评估和相关的健康影响评估。拟写的小说 在健康研究中评估空气污染暴露的范例将极大地改善暴露的沟通 对政策制定者和公众的健康影响估计的不确定性，恰好是对以下之一的回应 美国国立卫生研究院的优先研究领域。我们的工具可以很容易地扩展，并将有利于信息和 不同地点和全球范围以及其他环境的不确定性表征 曝光。

Project Summary Ambient air pollution is a global environmental threat, contributing to millions of deaths and hundreds of millions of disability-adjusted-life-years (DALYs) annually. However, the major limitation of air pollution health studies remains exposure assessment. Although there have been great advances in air pollution assessment and several sophisticated spatio-temporal models have been developed to predict daily air pollution levels at the residential addresses of study participants, the performance of these models varies in space and time. Even the best on average performing prediction model, however, will have limited predictive ability in certain space and time points. Furthermore, the uncertainty associated with use of a single prediction model has been consistently ignored in health studies, which could lead to invalid inferences of the health effect estimates, and inconsistent findings across studies. We propose to address this critical gap by developing a novel ensemble model framework for exposure assessment in air pollution health studies, integrating information across multiple existing prediction models. With this approach, we will for the first time be able to comprehensively quantify any inter- and intra-model uncertainty associated with ambient air pollution exposures. We will develop ensemble methods both for single- and multi-pollutant settings. We propose to apply the developed methods and fully propagate exposure uncertainty in health effect estimation using two nationwide open cohorts, mainly Medicare and Medicaid, as well as an open cohort of hospital admissions in New York State (Statewide Planning and Research Cooperative System, SPARCS). These datasets provide information on approximately all elderly, low-income and disabled Americans across the United States (Medicare and Medicaid, respectively), with residential information the zip-code level, as well as 98% of all hospitalizations in NY State, with information available at the residential address. Specifically we will assess the long- and short-term impact of air pollution exposure on mortality (Medicare and Medicaid), and cardiorespiratory morbidity (all three cohorts).We communicate the air pollution predictions, the spatio-temporal uncertainty of air pollution exposure assessment and related health effect estimates to the public and regulatory agencies. The proposed novel paradigm to assess air pollution exposures in health studies will greatly improve communication of exposure uncertainty in the health effect estimates both to policy makers and the public, exactly responding to one of NIH's priority research areas. Our tools can be easily extended and will benefit integration of information and uncertainty characterization at different locations and at a global scale, as well as for other environmental exposures.