Natural Language Question Understanding for Electronic Health Records

 描述(由申请人提供)：电子健康记录(EHR)中的患者信息，如实验室结果、药物和过去的病史，是医生决定患者护理的基础。它还帮助患者更好地理解和管理他们的护理。因此，有效地访问这些患者信息是至关重要的。访问数据最直观的方式之一是询问自然语言问题。在医学问题回答方面已经进行了大量的工作，但在EHR的问题回答方面所做的工作很少。自然语言问题可以用逻辑形式表示，这是一种标准的结构化知识表示技术。该项目建议采用自然语言电子病历问题，为医生和患者，并自动转换为逻辑形式。然后可以将逻辑表单转换为结构化查询，如EHR使用的查询。这种方法的一个主要障碍是缺乏包含用逻辑形式注释的问题的数据。这个项目假设可以手动注释一小部分问题，然后可以为每个注释的问题生成释义。由于释义是一项比逻辑形式注释更简单的任务，众包技术可以用来收集成千上万的问题释义。这个问题释义语料库将被用来建立一个能够识别EHR问题的逻辑结构的语义语法。为了确保一个健壮的、可概括的语法，现有的自然语言处理技术将被用来对问题进行预处理，简化其句法结构并抽象其医学概念。 为了开发这种方法，候选人柯克·罗伯茨博士需要在自然语言处理和生物医学信息学方面进行额外的培训和指导。这份申请NIH独立之路奖(K99/R00)的申请书描述了一项职业发展计划，该计划将使罗伯茨博士能够实现该项目的目标，并过渡到 独立研究人员。他将由领先的医学NLP研究员Dina Demner-Fushman博士指导，并由领先的电子病历和医疗信息学研究人员Clement McDonald博士共同指导。 该项目的具体目标是：(1)建立一个电子病历问题释义集合，其中每个原型问题都将有许多独特的释义。释义包含了表达相同逻辑形式的不同词汇和句法手段。(2)构建EHR疑问句的语义语法。然后，可以使用该语法将自然语言问题转换为逻辑形式。(3)实现了一个端到端的问题分析器，该分析器对EHR问题进行泛化以改进解析，使用语法将问题解析为逻辑形式，并将该逻辑形式转换为领先的结构化EHR查询格式。

 DESCRIPTION (provided by applicant): Patient information in the electronic health record (EHR) such as lab results, medications, and past medical history is the basis for physician decisions about patient care. It also helps patients better understand and manage their care. Efficient access to this patient information is thus essential. One of the most intuitive ways of accessing data is by asking natural language questions. A significant amount of work in medical question answering has been conducted, yet little work has been performed in question answering for EHRs. Natural language questions can be represented in logical forms, a standard structured knowledge representation technique. This project proposes to take natural language EHR questions, both for doctors and patients, and automatically convert them to a logical form. The logical forms can then be converted to a structured query such as those used by EHRs. A major obstacle to this approach is the lack of data containing questions annotated with logical forms. This project hypothesizes that a small set of questions can be manually annotated, and then paraphrases can be produced for each annotated question. Since paraphrasing is a simpler task than logical form annotation, crowd-sourcing techniques can be used to collect thousands of question paraphrases. This question paraphrase corpus will then be used to build a semantic grammar capable of recognizing the logical structure of EHR questions. To ensure a robust, generalizable grammar, existing NLP techniques will be used to pre-process questions, simplifying their syntactic structure and abstracting their medical concepts. In order to develop such a method, the candidate, Dr. Kirk Roberts, requires additional training and mentoring in natural language processing and biomedical informatics. This application for the NIH Pathway to Independence Award (K99/R00) describes a career development plan that will allow Dr. Roberts to achieve the goals of this project as well as transition to a career as an independent researcher. He will be mentored by Dr. Dina Demner-Fushman, a leading medical NLP researcher, and co-mentored by Dr. Clement McDonald, a leading EHR and medical informatics researcher. The specific aims of the project are: (1) Build a paraphrase collection of EHR questions, where each prototype question will have many unique paraphrases. The paraphrases encompass different lexical and syntactic means of conveying the same logical form. (2) Construct a semantic grammar for EHR questions. The grammar can then be used to convert a natural language question to a logical form. (3) Implement an end- to-end question analyzer that generalizes EHR questions for improved parsing, parses the question into a logical form using the grammar, and converts the logical form into a leading structured EHR query format.