CAREER: Machine Learning for Complex Health Data Analytics

健康和行为科学领域目前正在经历一场数据革命。2009年的《卫生信息技术促进经济和临床卫生法案》导致电子健康记录得到广泛采用，并出现了日益庞大的异类临床数据存储。与此同时，新兴的移动健康(MHealth)技术使得在非临床环境中收集越来越多的连续生理测量和行为自我报告数据成为可能。这样的数据来源有可能在对人类行为和健康的基本理解方面产生变革性的进展。它们还有可能显著增强众多应用，包括数据驱动的临床决策支持和持续的健康监测，这将提高医疗系统的效率，并促进向以患者为中心的个性化医疗的过渡。拟议的工作将解决临床和mHealth数据分析中的几个基本复杂性来源，使健康和行为科学的研究人员能够从这些数据来源中提取更多有用的知识。将开发的软件工具箱将立即在研究伙伴网络进行的研究中应用，并将广泛传播。综合教育计划包括开发一门创新的应用机器学习课程，该课程将提供与海量健康数据分析直接相关的云计算等主题的培训。外联计划涉及为代表不足的群体开发和开办以健康数据为主题的外联讲习班，以促进计算思维并扩大对计算的参与。从复杂数据中学习模型并应用这些模型提取有用知识的能力是机器学习研究的核心。这项提议寻求通过开发新的模型和算法来显著扩展机器学习的前沿，以应对复杂的健康数据分析带来的挑战。临床和移动健康数据复杂性的主要来源包括稀疏和不规则的采样、不完整、噪声、非平稳的时间动态、受试者之间的可变性、大容量、高速和异质性。在给定的数据源中存在这些因素中的一个或多个通常足以使当前的机器学习方法无效或完全不适用。这项研究的长期目标是开发和验证定制的机器学习模型和算法，以应对所有这些挑战。该建议的目标是开发解决以下具体问题的模型和算法：(1)如何从稀疏和不规则采样的临床时间序列数据中提取有用的知识？(2)在存在高水平噪声、受试者之间显著的可变性和不同的感知模式的情况下，如何从可穿戴式生理传感器数据中自动发现特征？(3)如何使生理时间序列事件检测算法的学习对通过有限可靠性和时间保真度的自我报告机制获得的事件标签具有健壮性。

The fields of health and behavioral science are currently undergoing a data revolution. The Health Information Technology for Economic and Clinical Health act of 2009 has resulted in the wide adoption of electronic health records and the emergence of increasingly vast stores of heterogeneous clinical data. Simultaneously, emerging mobile health (mHealth) technologies are enabling the collection of ever-larger volumes of continuous physiological measurements and behavioral self-report data in non-clinical settings. Such data sources have the potential to yield transformative advances in the fundamental understanding of human behavior and health. They also have the potential to significantly enhance numerous applications including data-driven clinical decision support and continuous health monitoring, which will lead to increased efficiency within the healthcare system and facilitate a transition to patient-centered, personalized care. The proposed work will address several fundamental sources of complexity in the analysis of both clinical and mHealth data, enabling researchers in health and behavioral science to extract more useful knowledge from these data sources. The software toolboxes that will be developed will have immediate applications in research conducted by a network of research partners, and will also be broadly disseminated. The integrated education plan includes the development of an innovative applied machine learning course that will provide training in topics like cloud-scale computing that are of direct relevance to massive health data analytics. The outreach plan involves developing and running a health data-themed outreach workshop for underrepresented groups to foster computational thinking and broaden participation in computing. The ability to learn models from complex data and apply those models to extract useful knowledge is at the core of machine learning research. This proposal seeks to significantly expand the frontiers of machine learning by developing new models and algorithms designed to meet the challenges posed by complex health data analysis. Key sources of complexity in clinical and mHealth data include sparse and irregular sampling, incompleteness, noise, non-stationary temporal dynamics, between-subjects variability, high volume, high velocity and heterogeneity. The presence of one or more of these factors in a given data source is often sufficient to render current machine learning methods ineffective or completely inapplicable. The long-term goal of this research is the development and validation of customized machine learning models and algorithms that can respond to all of these challenges. The objective of this proposal is to develop models and algorithms that address the following specific problems: (1) How can we extract useful knowledge from sparse and irregularly sampled clinical time series data? (2) How can we automate feature discovery from wearable physiological sensor data in the presence of high levels of noise, significant between subjects variability, and heterogeneous sensing modalities? (3) How can we make the learning of physiological time series event detection algorithms robust to event labels that are obtained through self-report mechanisms with limited reliability and temporal fidelity?