EAGER: Collaborative Research: Advanced Machine Learning for Prediction of Preterm Birth

美国每年花费超过260亿美元用于12-13%的早产婴儿的分娩和护理。由于早产是一个重大的公共卫生问题，对社会有着深远的影响，因此能够在怀孕期间识别有早产风险的妇女将具有极大的价值。以前的预测方法在很大程度上是不成功的，因为它们集中在有限数量的已知与早产相关的已充分描述的风险因素（例如，先前的早产，种族和感染），而不是结合多种因素。后一种方法对于理解早产的复杂病因是必要的。虽然识别个体PTB风险因素已经使人们深入了解了这个问题，并在某些情况下导致了成功的治疗，例如对先前早产的妇女使用黄体酮，但这对总体频率的影响有限，因为许多风险患者，例如第一次母亲（未经产），未经治疗。今天，仍然没有广泛测试的预测系统，结合众所周知的PTB因素，并在临床上有用。然而，全球意识到需要发现和整合早产的复杂病因，以预测处于风险中的妇女。在过去的几十年里，人们做出了巨大的努力来收集孕妇的大型策划数据集。以前对这些数据集的研究使用相对简单的生物统计方法，如相对风险评估来衡量因素与PTB之间的关联。然而，风险因素的研究是相互独立的，这并不能解释PTB的多因素复杂性。这个探索性项目旨在通过同时考虑所有因素来研究更先进的机器学习方法的价值，以开发更好的预测方法。在该项目的背景下获得的PTB数据汇集了母亲及其婴儿的电子健康记录（EHR）沿着精心策划的NIH数据。 数据中包含大量的结构化临床数据和非结构化自由文本，需要手动提取特征。该项目主要由PTB问题驱动，有两个主要目标：（1）提高异构数据标注的质量和聚合。研究人员的目标是通过研究人类注释者的手动特征提取过程来捕获临床笔记文本中记录的社会经济，心理和行为风险因素。最先进的方法依赖于注释者的专业知识和/或实例的难度，但忽略了由于疲劳、无聊或知识而导致的标记质量随时间的变化。为了改进注释，该项目将开发一种新的贝叶斯框架，用于非结构化数据的人工标记。 贝叶斯模型将嵌入一套完整的参数，包括每个类别的流行程度，实例的难度和过程中注释质量的变化。如果模型构建成功，那么开发的框架将取代ad-hoc注释，成为一个精心设计的过程，以产生高质量的注释。该框架将允许从临床文本中提取可靠的特征，用于设计PTB预测模型的后续分析。(2)为多个数据空间开发预测模型。为了利用所有现有的数据，该项目将调查在早产的背景下使用Vapnik的学习使用标记信息（LUPI）范式的价值。标记信息是可用于训练模型但不可用于测试示例的数据。该项目中的数据有两个潜在的特权信息空间，即临床记录和未来事件空间。 NICU数据是未来事件特权信息的示例，其仅可用于示例的子集（仅需要重症监护的早产儿留在NICU中）。结果表明，LUPI不仅诱导了更好的决策规则，它还提高了算法的收敛速度，因此需要更少的训练样本。由于PTB的发生率，在PTB预测的情况下，这是一个引人注目的属性。该项目将把LUPI扩展成一个强大而适用的框架，以处理两个空间的特权信息，同时开发样条生成内核，以管理LUPI的高计算成本。如果成功，这种概念验证预计将在特权信息可用的领域（如金融领域和许多其他医疗应用）产生高效且广泛适用的LUPI算法。该项目开发的软件、出版物和数据集将通过项目网站（http：//www1.ccls.columbia.edu/columbiansaf/CING/PTB/）向研究界公开提供。

The United States spends over 26 billion dollars per annum on the delivery and care of the 12-13% of infants who are born preterm. As preterm birth (PTB) is a major public health problem with profound implications on society, there would be extreme value in being able to identify women at risk of preterm birth during the course of their pregnancy. Previous predictive approaches have been largely unsuccessful since they have focused on a limited number of well described risk factors known to be correlated with preterm birth (e.g. prior preterm birth, race, and infection) and less on combining multiple factors. The latter approach is necessary to understand the complex etiologies of preterm birth. While identifying individual PTB risk factors has brought insight into the problem and has led in some cases to successful treatments such as progesterone for women with a previous preterm birth, this has only a limited impact on the overall frequency since many at risk patients, such as first time mothers (nulliparous), go untreated. Today, there is still no widely tested prediction system that combines well-known PTB factors and is clinically useful. There is, however, a global awareness of the need to discover and integrate the complex etiologies of prematurity in order to predict women at risk. Significant efforts have been made in the last couple of decades to collect large curated datasets of pregnant women. Previous studies on these datasets used relatively straightforward biostatitistical methodologies such as relative risk assessments to measure associations between factors and PTB. However, risk factors are studied independently of each other, which does not account for the multifactorial complexity of PTB. This exploratory project aims to investigate the value of more advanced machine learning methods by simultaneously considering all the factors, to develop better predictive methods. The PTB data acquired in the context of this project brings together Electronic Health Records (EHRs) for mothers and their babies along with well-curated NIH data. The data is rich with structured clinical data and unstructured free text that require manual feature extraction. This project, largely motivated by the PTB problem, has two main goals: (1) Improving the quality and aggregation the annotations for heterogeneous data. The researchers aim to capture socioeconomic, psychological and behavioral risk factors documented in the text of clinical notes via studying the process of manual feature extraction by human annotators. State-of-the-art methods either rely on the expertise of the annotator and/or the difficulty of the instance but ignore the variability in the quality of labeling over time due to fatigue, boredom, or knowledge. To improve the annotations, the project will develop a novel Bayesian framework for human labeling of unstructured data. The Bayesian model will embed a complete set of parameters including the prevalence of each class, difficulty of the instance and variability in the quality of annotation during the process. If the model construction is successful, then the developed framework will replace ad-hoc heuristics into a well-designed process for producing high quality annotations. This framework would allow extracting reliable features from the clinical text for subsequent analyses in devising PTB prediction models.(2) Developing predictive models for multiple data spaces. To leverage all of the existing data, the project will investigate the value of using Vapnik's paradigm of Learning Using Privileged Information (LUPI) in the context of preterm birth. Privileged information is a data that is available for training models but is not available for test examples. Data in this project come with two potential privileged information spaces namely the clinical notes and the space of future events. NICU data is an example of future event privileged information, which is only available for a subset of the examples (only premature babies requiring intensive care stay in the NICU). It has been shown that LUPI not only induces a better decision rule, it also increases the rate of convergence of the algorithm, hence requiring fewer training examples. This is a compelling property in the case of PTB prediction because of the rate of PTB. The project will extend LUPI into a powerful and applicable framework to handle the two spaces of privileged information, while developing spline-generating kernels, to manage LUPI's high computational cost. If successful, this proof-of-concept is expected to yield efficient and widely applicable LUPI algorithms in domains where privileged information is available, such as the financial domain and many other medical applications. The developed software, publications and datasets resulting from this project will be made publicly available to the research community through the project website (http://www1.ccls.columbia.edu/~ansaf/CING/PTB/).