NARURAL LANGUAGE PROCESSING OF MEDICAL REPORTS

医疗报告包含大量描述病人健康状况的信息。然而，这些信息中有很大一部分是以自由文本的形式非结构化的。在这种形式下，信息很难搜索、分类、分析、汇总和呈现。我们提出了一种新的自然语言处理方法，试图从医学自由文本报告中自动提取重要概念。句子分析算法基于一种“场论”，该理论的动机是多粒子系统如何形成附件，正如物理学中所解释的那样。自然语言处理的这种观点试图将单词视为主动实体，而不是被动的数据元素。给出了一个如何以及为什么会出现单词附加的模型。该模型根据表征单词在句子分析的给定状态下的稳定性的能量场和涉及共振现象的信号处理模型来提供解释。我们的假设是：1)系统可以从医疗报告中提取信息；2)结果可以用规范化的形式表示出来；3)系统可以扩展到医学的许多领域。我们系统的测试和评估将根据特定放射学领域的报告进行。我们将探索NLP系统在儿科病理和泌尿外科领域的适应性。除了技术措施外，还将从最终用户的角度对NLP进行评估，包括选择适当的成像协议、SNOMED编码以及从放射学教学案例中检索信息。

Medical reports contain a great deal of information that characterizes a patient's medical condition. A large percent of this information, however, is unstructured in the form of free text. In this form, the information is difficult to search, sort, analyze, summarize and present. We present a new method of natural language processing which attempts to automatically extract the important concepts from medical free text reports. The sentence parsing algorithm is based on a "field theory" that is motivated by how a multi-particle system forms attachments as explained in physics. This view of natural language processing attempts to view words as active entities rather than passive data elements. A model for how and why word attachments occur is presented. The model presents explanations in terms of both energy fields that characterize the stability of a word at a given state of the sentence parse and in terms of a signal processing model involving a resonance phenomenon. Our hypotheses are that: 1) the system can extract information from medical reports; 2) the results can be represented in a canonical form with respect to meaning; 3) the system is extendable to many domains of medicine. The testing and evaluation of our system will be performed on reports from the specific domains of radiology. We will explore the adaptability of the NLP system to the domains of pediatric pathology and urology. In addition to technical measures, NLP will be evaluated from the end-user perspective including the selection of appropriate imaging protocols, SNOMED coding, and information retrieval from radiology teaching cases.