Built Environment Assessment through Computer visiON (BEACON): Applying Deep Learning to Street-Level and Satellite Images to Estimate Built Environment Effects on Cardiovascular Health

项目摘要 超过80%的美国人口居住在城市地区，而建筑环境-建筑物，街道， 我们生活的绿色空间--可能通过促进或限制体力活动而导致心血管疾病（CVD 和体重增加，并通过影响暴露于环境因素，如空气污染，极端 温度和噪音。建筑环境和CVD的证据主要来自横截面 非特异性暴露评估研究。开发精确、随时间变化和个性化的曝光 度量是必要的，以建立建筑环境和CVD之间的因果关系，这是至关重要的 为政策相关的、可操作的干预措施提供信息。现在可以估计这样的暴露度量， 使用深度学习计算机视觉方法（一类机器学习）的前瞻性队列研究量表 算法，可以准确地处理图像，结合时变全国街道级图像，高 高分辨率卫星数据和新型移动的保健技术。我们建议确定建筑物的影响 环境对CVD健康行为和CVD发病率的影响 从深度学习算法，并将这些暴露措施应用于参与者的时间活动数据， 来自护士健康研究3（N=500）的全球定位系统（GPS）数据，以及地理编码的住宅 来自全国护士健康研究、护士健康研究II和健康专业人员随访的地址 研究前瞻性队列（N= 288，000）。我们将通过充分利用 学习算法应用于全国谷歌街景图像（2007-2020年）和高分辨率陆地卫星 卫星数据（1986-2020年），以创建自然环境的精细尺度、随时间变化的建筑环境指标 (e.g.,树），物理环境（例如，人行道），感知（例如，安全）和城市形态（例如，紧凑 高层）。我们将使用一系列创新的分析方法来确定建筑环境的影响 CVD相关健康行为和CVD发病率在不同的时间范围。首先，我们将这些 收集了GPS和身体活动的分钟级数据的参与者的时间活动模式的指标 从智能手机和消费者可穿戴设备，以量化如何分钟级暴露于建成 环境与CVD健康行为有关。接下来，我们将把新颖的建筑环境指标应用于住宅 解决参与者的历史，以估计自我报告的CVD健康行为如何改变后， 住宅建筑环境的变化。最后，我们将研究长期累积 住宅暴露于建筑环境和CVD发病率超过34年的随访。我们的工作将使 我们从前所未有的角度在大型前瞻性队列中测量建筑环境暴露， 阐明建筑环境和CVD健康行为之间的潜在因果关系，并更好地 指明CVD发病途径。最终，我们的工作将产生可操作的见解，以指导土地使用政策 和城市规划战略，以设计优化心血管健康的城市。

PROJECT SUMMARY Over 80% of the US population resides in urban areas, and the built environment—the buildings, streets, and green spaces in which we live—may drive cardiovascular disease (CVD) by promoting or limiting physical activity and weight gain, and by influencing exposures to environmental factors, such as air pollution, extreme temperatures, and noise. Evidence for the built environment and CVD has been dominated by cross-sectional studies with nonspecific exposure assessment. Developing precise, time-varying, and personalized exposure metrics is necessary to establish causal relationships between the built environment and CVD, which are crucial to informing policy-relevant, actionable interventions. It is now possible to estimate such exposure metrics at scale in prospective cohort studies using deep learning computer vision methods, a class of machine learning algorithms that can accurately process images, combined with time-varying nationwide street-level imagery, high resolution satellite data, and novel mobile health technologies. We propose to identify the influence of the built environment on CVD health behaviors and CVD incidence by developing built environment exposure measures from deep learning algorithms, and to apply these exposure measures to time-activity data in participants with global positioning systems (GPS) data from the Nurses’ Health Study 3 (N=500), and to geocoded residential addresses from nationwide Nurses’ Health Study, Nurses’ Health Study II, and Health Professionals Follow-up Study prospective cohorts (N=288,000). We will create built environment exposure measures by leveraging deep learning algorithms applied to nationwide Google Street View imagery (2007-2020) and high-resolution Landsat satellite data (1986-2020) to create fine-scale, time-varying built environment metrics of the natural environment (e.g., trees), physical environment (e.g., sidewalks), perceptions (e.g., safety), and urban form (e.g., compact high-rise). We will use a mix of innovative analytical approaches to determine the effect of the built environment on CVD-related health behaviors and CVD incidence across different time horizons. First, we will append these metrics to time-activity patterns of participants who have collected minute-level data on GPS and physical activity from smartphones and consumer wearable devices to quantify how minute-level exposure to the built environment is related to CVD health behaviors. Next, we will apply novel built environment metrics to residential address histories of participants to estimate how self-reported CVD health behaviors change after their residential built environment changes. Last, we will examine the association between long-term cumulative residential exposure to the built environment and CVD incidence over 34 years of follow-up. Our work will enable us to measure built environment exposure from unprecedented perspectives in large prospective cohorts, to elucidate potential causal relationships between the built environment and CVD health behaviors, and to better specify pathways to CVD incidence. Ultimately, our work will yield actionable insights to guide land use policy and urban planning strategies to design cities that optimize cardiovascular health.