CAREER: Learning from When, Where and by Whom Data is Generated for Advancing Public Health Studies

通过对日常生活中的生活方式、环境和社会因素进行强有力和高粒度的测量来改善疾病预防，将通过实现精确和重点突出的主动干预来改善医疗保健。这将极大地改变这个国家的医疗保健模式，并显著降低成本和疾病，而不仅仅是被动地关注疾病诊断和治疗。公共卫生是对这些日常生活因素和预防措施的研究。新的个人生成数据（PGD）来自互联网和移动数据源，如移动健康、社交媒体、可穿戴设备和智能手机应用程序的数据，提供了前所未有的机会，可以提供日常生活中生活方式、环境和社会因素的亚日常、本地、社区水平的测量。然而，这些数据对公共卫生工作的影响尚未得到充分认识。在某种程度上，这是因为现有的PGD研究工作往往侧重于孤立地处理数据内容，而没有考虑人类数据共享模式，即谁提供数据、何时提供数据以及从哪里提供数据。通过考虑这些属性，本项目旨在提高从PGD中提取的度量的有效性和可靠性，并提高对高粒度健康风险和结果的理解。该项目还将为PGD和公共卫生背景下的公共卫生从业人员、学生和社区成员提供一个高度集成的研究和教育计划：(1)通过基于问题的学习课程，使学生在当今的工作环境中使用计算机科学；(2)增加高中生在现实世界中接触计算机科学的机会，重点关注计算机科学的应用；(3)在公共卫生和一般社区传播计算机科学的科学理解。同时，这项工作将通过方法开发和让不同的社区成员和面向社区的专业人员接触数据挖掘和机器学习的效用，改善计算机科学和公共卫生实践和研究。该项目的目标是开发新的机器学习方法，其动机是需要改进公共卫生领域的数据管理和分析。该研究解决了由于人类数据共享模式而导致的关键统计和计算挑战。这些挑战为卫生信息学和机器学习做出贡献提供了机会，方法是通过从个人生成的数据中学习，结合“何时、何地、由谁”生成数据，改进预测工作。利用这些信息作为额外的信号，本项目探索：(1)通过考虑人类数据共享模式的特征来推断时间模式（motif）；(2)通过将内容与地理信息结合使用，发现人类内容在空间上产生的噪声、稀疏和不一致的潜在空间表征；(3)在没有标签的数据中进行预测，使用来自不同领域的相同任务的数据，通过包含每个数据集中生成数据的群体的属性。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Improving disease prevention through robust and high-granularity measures of lifestyle, environmental and social factors from daily life will improve healthcare by enabling precise and focused proactive interventions. This will dramatically change the healthcare paradigm in this country and significantly reduce costs and illnesses, more so than a solely reactive focus on disease diagnosis and treatment. Public health is the study of these daily life factors and prevention efforts. New person-generated data (PGD) from Internet and mobile data sources, such as mHealth, social media, wearables, and data from smartphone apps, offer unprecedented opportunity to provide sub-daily, as well as local, neighborhood-level measures of lifestyle, environmental and social factors from daily life. However, the impact of this data has yet to be fully realized for public health efforts. In part, this is because existing research efforts on PGD often focus on processing the content of data in isolation, and do not consider human data sharing patterns, that is, who contributes the data, when it is contributed and from where it is contributed. By accounting for these attributes, this project aims to improve the validity and reliability of measures extracted from PGD and enable improved understanding of high-granularity health risks and outcomes. The project will also provide a highly-integrated research and educational program for public health practitioners, students, and community members in the context of PGD and public health by: (1) preparing students to use computer science in today's job landscape via a problem-based learning class; (2) increasing high-school students' exposure to computer science in the real-world with a focus on applications of computer science; and (3) disseminating scientific understanding of computer science in the public health and general community. In conjunction, this work will improve both computer science and public health practice and research through method development and exposure of diverse community members and community-oriented professionals to the utility of data mining and machine learning. The goal of this project is to develop new machine learning approaches motivated by the need to improve data management and analysis in the public health domain. The research addresses critical statistical and computational challenges due to human data sharing patterns. These challenges represent an opportunity for contributions to health informatics and machine learning by improving prediction efforts through learning from person-generated data in combination with "when, where and by whom" the data is generated. Using this information as an additional signal, this project explores: (1) inference of temporal patterns (motifs) by accounting for characteristic human data sharing patterns; (2) discovery of underlying latent spatial representation of content from humans that is noisy, sparse and inconsistently generated over space by using content jointly with geographic information; and (3) prediction in data without labels using data for the same task but from a different domain by including attributes of the population generating the data in each dataset.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.