Semi-structured Information Retrieval in Clinical Text for Cohort Identification

描述（由申请人提供）：自然语言处理（NLP）技术已经显示出从电子健康记录（EHRs）的自由文本中提取数据的希望，但研究一致发现，该技术不容易在应用程序设置中推广。不幸的是，将NLP应用于实际用例的大多数焦点仍然停留在单一的、定义良好的应用程序设置的范例上，因此，对不可见用例的通用性仍然没有得到解决。我们建议通过采用信息检索（IR）的视角，以患者水平的队列识别为目标，明确地解释未见过的应用程序设置。为此，我们引入了分层语言模型，这是一个IR框架，可以重用nlp生成的工件。我们的长期目标是通过提供强大的、以用户为中心的工具来加速对病人健康和疾病的调查，这些工具是处理、检索和利用电子病历免费文本所必需的。本提案的主要目标是从梅奥诊所和OHSU的临床文献中准确地检索特别的、现实的队列，建立方法、资源和评估患者水平的IR。我们假设队列识别可以通过一种新的IR框架：分层语言模型以一种可推广的方式解决。我们将通过四个具体目标来检验这一假设。在目标1中，我们将使医学NLP工件在我们的分层语言IR框架中可搜索。这包括存储和索引NLP工件，以及使用统计语言模型来检索基于文本及其相关NLP工件的文档。在目标2中，我们处理临时队列识别的实际设置，转移到患者级（而不是文档级）IR。为了准确地处理患者队列，其中合格的证据可能分布在多个文件中，我们将开发和实施患者级检索模型，该模型考虑了跨文件关系和事件的时间组合。在目标3中，我们将使用来自两个站点的EHR数据构建并行IR测试集合；由多个队列编写的一组不同的队列查询

DESCRIPTION (provided by applicant): Natural Language Processing (NLP) techniques have shown promise for extracting data from the free text of electronic health records (EHRs), but studies have consistently found that techniques do not readily generalize across application settings. Unfortunately, most of the focus in applying NLP to real use cases has remained on a paradigm of single, well-defined application settings, so that generalizability to unseen use cases remains implicitly unaddressed. We propose to explicitly account for unseen application settings by adopting an information retrieval (IR) perspective with the objective of patient-level cohort identification. To do so, we introduce layered language models, an IR framework that enables the reuse of NLP-produced artifacts. Our long term goal is to accelerate investigations of patient health and disease by providing robust, user- centric tools that are necessary to process, retrieve, and utilize the free text of EHRs. The main goal of this proposal is to accurately retrieve ad hoc, realistic cohorts from clinical text at Mayo Clinic and OHSU, establishing methods, resources, and evaluation for patient-level IR. We hypothesize that cohort identification can be addressed in a generalizable fashion by a new IR framework: layered language models. We will test this hypothesis through four specific aims. In Aim 1, we will make medical NLP artifacts searchable in our layered language IR framework. This involves storing and indexing the NLP artifacts, as well as using statistical language models to retrieve documents based on text and its associated NLP artifacts. In Aim 2, we deal with the practical setting of ad hoc cohort identification, moving to patient-level (rather than document-level) IR. To accurately handle patient cohorts in which qualifying evidence may be spread over multiple documents, we will develop and implement patient-level retrieval models that account for cross- document relational and temporal combinations of events. In Aim 3, we will construct parallel IR test collections using EHR data from two sites; a diverse set of cohort queries written by multiple people toward various clinical or epidemiological ends; and assessments of which patients are relevant to which queries at both sites. Finally, in Aim 4, we refine and evaluate patient-level layered language IR on the ad hoc cohort identification task, making comparisons across the users, queries, optimization metrics, and institutions. We will draw additional extrinsic comparisons with pre-existing techniques, e.g., for cohorts from the Electronic Medical Records and Genonmics network. The expected outcomes of the proposed work are: (i) An open-source cohort identification tool, usable by clinicians and epidemiologists, that makes principled use of NLP artifacts for unseen queries; ii) A parallel test collection for cohort identification, includig two intra-institutional document collections, diverse test topics and user-produced text queries, and patient-level judgments of relevance to each query; and (iii) Validation of the reusability of medical NLP via the task of retrieving patient cohorts.