III: Small: Discovering Complex Anomalous Mappings

美国医学研究所（Institute of Medicine） 2000年的一份报告显示，美国每年有10万名患者死亡，原因是医疗保健专业人员在处理大量数据的同时，在时间压力下做出了不理想的治疗决定，尤其是在手术室和重症监护病房。纠正措施和新规定尚未带来改善。在发达国家，大约6%的经济用于基础设施的维护，大约50%的原始购置成本用于设备的维护。例如，机队的管理人员往往无法维持所需的设备可用性水平，这是由于意想不到的，但在数据中可以识别的维护和后勤危机：只有大约2/3的美国军用飞机可以同时飞行。快速解决方案通常会导致数亿本可避免的费用。这两个例子，以及许多其他社会、经济和科学上重要的人类活动领域，涉及大量的多流数据，这些数据可能携带有助于减轻某些逆境的信息。现有的研究成果产生了从单个数据源中提取有用信息的算法。然而，通过利用数据流之间的关系，可以实现实质性的新好处。这项研究计划将全面而务实地探索这一机会，并影响医疗保健从业人员、设备管理人员以及其他领域的用户，只要有多种确凿证据可用，它也将使学生和学员以及整个科学界受益。该研究项目将开发并广泛评估新的算法，以识别大型、多元、数字和符号、潜在稀疏数据的多个和不同流之间的信息相关性。这些算法将识别特征子集和数据记录，遵循不同的跨流关系模式，在静态和时间设置中实现描述性和预测性分析，并允许稳健的预测。拟议的工作将建立在先前对多维数据单流中复杂异常模式检测的努力基础上，并将其扩展到多个数据源的跨流分析。预期结果将允许，例如，检测在重症监护病人床边测量的生命体征与他们的治疗或药物记录之间关系的变化模式，以及结果。这些模式可能表明患者对治疗的反应不标准，或表明出现了健康危机。从更广泛的角度来看，这项工作将大大扩展当前跨流分析技术的能力，如典型相关分析，它将开发高斯过程框架的新变体，以模拟流间动力学，产生符号变量流之间相关性的新的信息理论模型（扩展到处理混合数字和符号特征的数据集）。它将为识别跨流关系的多模态结构提供一个框架，包括析取和合取-析取模式。它将采取极具挑战性的智力努力，同时以社会重要性的显著效益为目标，主要影响示范领域是床边信息学和设备健康管理。新算法的最终软件实现及其使用的说明性示例将与一般研究社区共享。大部分拟议的工作将由研究生和医学研究员完成，他们将立即在其职业生涯中使用所获得的知识。这些项目将把结果包括在外联活动及其培训和教学课程材料中。

A 2000 report by the Institute of Medicine indicated that 100,000 patients die in the U.S. each year due to suboptimal treatment decisions made by healthcare professionals who operate under pressure of time while processing overwhelming amounts of data, particularly in the operating rooms and intensive care units. Corrective measures and new regulations have not led to improvements yet. In developed countries, about 6% of the economy is spent on upkeep of infrastructure, and about 50% of the original acquisition cost is spent on maintenance of equipment. Managers of, e.g., fleets of aircraft, are too often unable to maintain the required levels of equipment availability due to unexpected, but identifiable in data, crises in maintenance and logistics: only about 2/3 of the U.S. military aircraft can be flown at once. Expediting solutions often causes hundreds of millions in avoidable expenses. These two examples, as well as many other societally, economically, and scientifically important domains of human activity, involve large amounts of multi-stream data, which may carry information helpful in mitigating some of the adversities. Existing research efforts produce algorithms that extract useful information from individual sources of data. Substantial new benefits, however, could be realized by exploiting relationships between streams of data. This research program will comprehensively and pragmatically explore that opportunity, and impact communities of healthcare practitioners, equipment managers, as well as users in other domains wherever multiple streams of corroborative evidence are available, it will also benefit students and trainees, and the scientific community at large.This research project will develop and extensively evaluate new algorithms to identify informative correlations between multiple and diverse streams of large, multivariate, numeric and symbolic, potentially sparse data. These algorithms will identify subsets of features and records of data that follow distinct cross-stream relationship patterns, enable descriptive and predictive analytics in static and temporal settings, and allow robust forecasts. The proposed work will build on prior efforts towards detection of complex anomalous patterns in single streams of multidimensional data, and expand it towards cross-stream analysis of multiple data sources. Expected results will allow, e.g., detection of patterns of change in relationships between vital signs measured at the bedside of intensive care patients and their records of treatment or medication, and outcomes. These patterns may be indicative of non-standard responses of a patient to a treatment, or signal emergence of a health crisis. In a broader perspective, this effort will substantially expand capabilities of current techniques of cross-stream analytics such as Canonical Correlation Analysis, it will develop a new variant of Gaussian Processes framework to model inter-stream dynamics, produce new information-theoretical modeling of correlations between streams of symbolic variables (with an extension to handle datasets with a mix of numeric and symbolic features), and it will provide a framework for identifying multimodal structures of cross-stream relationships, including disjunctive and conjunctive-disjunctive patterns. It will take on highly challenging intellectual endeavors while aiming for significant benefits of societal importance, with the primary impact demonstration areas in bed-side informatics and equipment health management. Resulting software implementations of the new algorithms and illustrative examples of their use will be shared with the general research community. The bulk of the proposed work will be performed by graduate students and medical fellows who will immediately use the acquired knowledge in their careers. The PIs will include the results in outreach activities and in their training and teaching course materials.