Adapting Natural Language Processing Tools for Biosurveillance

描述（由申请人提供）： 及早发现疾病暴发可降低患者发病率和死亡率，并尽量减少疾病的传播。早期检测需要在患者病程早期对患者症状进行准确分类。一种方法是生物监测，即在疾病过程的早期捕获电子症状数据，并分析可能表明需要公共卫生系统调查和应对的疫情的信号。急诊科（艾德）病历对于生物监测特别有用，因为它们可以及时以电子方式提供。监测系统中使用的艾德数据元素包括主诉（患者主要症状的简要描述）和分诊护士记录（也称为当前病史）。主诉是最广泛使用的艾德数据元素，因为大多数ED都以电子方式记录。一项研究表明，增加分诊说明提高了生物监测病例检测的灵敏度。增加的敏感性是因为分类说明增加了可用的数据量：而不是主诉中的一个症状（例如，发烧），分诊记录可能包含多种症状（例如，“发烧、咳嗽和呼吸急促12小时”）。然而，艾德主诉和分诊记录中自由文本数据的广泛变异性阻碍了监测工作。它们通常包括拼写错误、缩写、首字母缩略词和其他词汇和语义变体，这些变体难以归类为症状群（例如，发热，体温104，FVR，发热）。需要工具来解决艾德数据中症状术语的词汇和语义变化，以改善生物监测。自然语言处理工具已被证明有助于从更结构化的临床数据（如放射学报告）中提取概念，但这些技术在自由文本艾德分诊笔记中的应用有限。项目团队开发了紧急医疗文本处理器（EMT-P）来预处理主诉。EMT-P清理和规范化简短的主诉条目，然后提取标准化的概念，但在目前的状态下，它不足以预处理更长，更复杂的文本段落，如分诊笔记。这一拟议项目将进一步加强生物监测，方法是调整EMT-P和其他统计和经典自然语言处理工具，开发一个从生物监测分类说明中提取概念的系统。

DESCRIPTION (provided by applicant): Early detection of disease outbreaks can decrease patient morbidity and mortality and minimize the spread of diseases. Early detection requires accurate classification of patient symptoms early in the course of their illness. One approach is biosurveillance, in which electronic symptom data are captured early in the course of illness, and analyzed for signals that might indicate an outbreak requiring investigation and response by the public health system. Emergency department (ED) patient records are particularly useful for biosurveillance, given their timely, electronic availability. ED data elements used in surveillance systems include the chief complaint (a brief description of the patient's primary symptom(s)), and triage nurses' note (also known as history of present illness).The chief complaint is the most widely used ED data element, because it is recorded electronically by the majority of EDs. One study showed that adding triage notes increased the sensitivity of biosurveillance case detection. The increased sensitivity is because the triage note increases the amount of data available: instead of one symptom in a chief complaint (e.g., fever), triage notes may contain multiple symptoms (e.g., "fever, cough & shortness of breath for 12 hours"). Surveillance efforts are hampered, however, by the wide variability of free text data in ED chief complaints and triage notes. They often include misspellings, abbreviations, acronyms and other lexical and semantic variants that are difficult to group into symptom clusters (e.g., fever, temp 104, fvr, febrile). Tools are needed to address the lexical and semantic variation in symptom terms in ED data in order to improve biosurveillance. Natural language processing tools have been shown to facilitate concept extraction from more structured clinical data such as radiology reports, but there has been limited application of these techniques to free text ED triage notes. The project team developed the Emergency Medical Text Processor (EMT-P) to preprocess the chief complaint. EMT-P cleans and normalizes brief chief complaint entries and then extracts standardized concepts, but it is not sufficient in its current state to preprocess longer, more complex text passages such as triage notes. This proposed project will further strengthen biosurveillance by adapting EMT-P and other statistical and classical natural language processing tools to develop a system that extracts concepts from triage notes for biosurveillance.