Bridging the Semantic Gap Between Research Eligibility Criteria and Clinical Data

描述(由申请人提供)： 我们的长期目标是通过加强临床数据的使用来支持临床研究决策，从而扩大临床研究的范围和效率。该提案旨在通过开发新的语义对齐框架来改进电子健康记录(EHR)的使用，以自动化临床试验资格筛选。临床试验研究是将基础生物医学科学的突破转化为有利于临床实践和人类健康的知识的重要一步。然而，一个重大障碍是确定合格的参与者。在所有临床试验中，有86%的患者招募延迟了一到六个月，13%的患者招募延迟了六个月以上。入学延迟代价高昂。在最近的一次大型多中心试验中，每个参与者都花费了大约86.8个工作时间和1000多美元。无效的登记也会产生巨大的社会成本，高达60%的患者可能会错过被识别的机会。电子病历系统的广泛部署为解决这一问题创造了前所未有的机会，因为电子病历系统包含关于潜在参与者的丰富信息来源。然而，要将符合条件的标准(如“肾功能不全”)与临床数据(如“80岁白人女性患者的血清肌酐=1.0毫克/分升”)相匹配，往往是一种知识密集型、耗时和低效的手动程序。这一持久的挑战部分是由于抽象和模糊的资格标准与高度具体的临床数据表现之间的脱节造成的；我们称之为语义鸿沟。尽管早期致力于基于计算机的临床指南和方案的工作，但在支持概念与其在患者表型(如体征和症状)中的表现之间的自动匹配方面投入的努力有限。 我们假设我们可以刻画语义鸿沟，并设计一个基于知识的、自然语言处理辅助的语义对齐框架来弥合语义鸿沟。因此，我们的具体目标是：(1)调查临床试验资格标准和临床数据之间的语义差距；(2)设计基于概念的、可计算的资格标准知识表示；(3)设计一个连接资格标准知识库和临床数据仓库的语义匹配框架，以生成用于资格识别的语义查询；以及(4)评估语义匹配框架的实用性。 这项研究的新颖和独特之处在于：(1)以前没有关于资格标准和临床数据之间的语义差距的研究；(2)我们首次设计了一个语义对齐框架，将资格标准与临床数据自动匹配。由哥伦比亚大学生物医学信息学系和加州大学旧金山分校普通医学部的专业知识组成的研究团队处于开展这项研究的独特地位，考虑到团队的经验(医学知识表示、自然语言处理、受控临床术语、基于本体的语义推理、数据挖掘、统计学、健康数据组织、语义协调和临床试验)，200万患者13年数据的存储库的可用性，以及名为MedLEE的自然语言处理器的可用性，可以将数百万叙述性报告转换为丰富编码的临床数据。

DESCRIPTION (provided by applicant): Our long-term objective is to enlarge the scope and efficiency of clinical research through enhanced use of clinical data to support clinical research decisions. This proposal aims to improve the use of electronic health records (EHR) to automate clinical trials eligibility screening by developing a new semantic alignment framework. Clinical trials research is an important step for translating breakthroughs in basic biomedical sciences into knowledge that will benefit clinical practice and human health. However, a significant obstacle is identifying eligible participants. Eighty-six percent of all clinical trials are delayed in patient recruitment for from one to six months and 13% are delayed by more than six months. Enrollment delay is expensive. In a recent large, multi-center trial, about 86.8 staff hours and more than $1000 was spent to enroll each participant. Ineffective enrollment also produces a big social cost in that up to 60% of patients can miss being identified. The broad deployment of EHR systems has created unprecedented opportunities to solve the problem because EHR systems contain a rich source of information about potential participants. However, it is often a knowledge-intensive, time-consuming, and inefficient manual procedure to match eligibility criteria such as "renal in- sufficiency" to clinical data such as "serum creatinine = 1.0 mg/dl for an 80-year old white female patient." This enduring challenge is partly caused by the disconnection between abstract and ambiguous eligibility criteria and highly specific clinical data manifestations; we call this a semantic gap. Despite earlier work on computer-based clinical guidelines and protocols, limited effort has been devoted to support automatic matching between concepts and their manifestations in patient phenotypes such as signs and symptoms. We hypothesize that we can characterize the semantic gap and design a knowledge-based, natural-language processing assisted semantic alignment framework to bridge the semantic gap. Therefore, our specific aims are: (1) to investigate the semantic gap between clinical trials eligibility criteria and clinical data; (2) to design a concept-based, computable knowledge representation for eligibility criteria; (3) to design a semantic alignment framework linking an eligibility criteria knowledge base and a clinical data warehouse to generate semantic queries for eligibility identification; and (4) to evaluate the utility of the semantic alignment framework. This research is novel and unique in that (1) there are no prior studies about the semantic gap between eligibility criteria and clinical data; and (2) for the first time, we design a semantic alignment framework to automatically match eligibility criteria to clinical data. The research team comprising expertise from the Department of Biomedical Informatics at Columbia University and the Division of General Medicine from UCSF are uniquely positioned to carry out this research, given the experience of the team (medical knowledge representation, natural language processing, controlled clinical terminology, ontology-based semantic reasoning, data mining, statistics, health data organization, semantic harmonization, and clinical trials), the availability of a repository of 13 years of data on 2 million patients, and the availability of a natural language processor called MedLEE to convert millions of narrative reports into richly coded clinical data.