Using NLP to Extract Clinically Important Recommendations from Radiology Reports

摘要 在发现异常时传达临床上重要的随访建议 成像研究很容易出错。缺乏识别和跟踪放射学随访的自动化系统 建议是确保患者及时随访的重要障碍，特别是对于非急性但 潜在的生命危险和意外的发现。这项提议的主要目标是发展一种自然的 从自由文本中提取临床重要推荐信息的语言处理(NLP)系统 放射学报告。每份放射学报告将在结构、句法和语义层面进行预处理，以 生成将用于提取包含推荐的句子边界的特征 信息以及推荐原因、请求的成像测试和推荐的详细信息 时间框架。我们将使用一个大型自由文本放射学报告语料库，这些报告由多种医疗模式混合表示 (例如，射线照相、计算机断层扫描、超声波和磁共振成像(MRI)) 不同的机构。使用此数据集，我们将实现以下具体目标：目标1.创建多个 为临床重要推荐信息注释的机构放射学报告语料库；目的2。 开发一种新的NLP系统来提取放射学报告中的临床重要建议。这个 拟议的研究具有创新性，因为它将产生一种新的文本处理方法，可用于标记 以可视和电子方式报告，以便可以启动单独的工作流程以减少机会 报告中建议的必要调查或干预措施被临床医生遗漏。建议数 一套工具将作为开放源码工具分发给生物医学信息学社区。

Abstract Communication of clinically important follow-up recommendations when abnormalities are identified on imaging studies is prone to error. The absence of an automated system to identify and track radiology follow-up recommendations is an important barrier to ensuring timely follow-up of patients, especially for non-acute but potentially life threatening and unexpected findings. The primary goal of this proposal is to develop a Natural Language Processing (NLP) system to extract clinically important recommendation information from free-text radiology reports. Each radiology report will be preprocessed at the structural, syntactic, and semantic level to generate features that will be used to extract the boundaries of sentences that include recommendation information as well as the details of reason for recommendation, requested imaging test, and recommendation time frame. We will use a large corpus of free-text radiology reports represented by a mixture of modalities (e.g., radiography, computed tomography, ultrasound, and magnetic resonance imaging (MRI)) from three different institutions. Using this dataset we will perform the following specific aims: Aim 1. Create a multi- institutional radiology report corpus annotated for clinically important recommendation information; Aim 2. Develop a novel NLP system to extract clinically important recommendations in radiology reports. The proposed research is innovative because it will generate a new text processing approach that can be used to flag reports visually and electronically so that separate workflow processes can be initiated to reduce the chance that necessary investigations or interventions suggested in the report are missed by clinicians. The proposed set of tools will be disseminated to the biomedical informatics community as open source tools.