SBIR Phase I: A tool to automate a narrative patient summary of the medical chart for outpatient physicians

小型企业创新研究(SBIR)第一阶段项目的更广泛影响/商业潜力是开发一种支持机器学习的病历摘要工具，旨在提供可帮助医生进行患者护理的叙述性摘要。医生平均只花3分钟查看患者的医疗记录，在此期间，他们必须解释非结构化电子健康记录(EHR)，这可能会使医生难以识别对患者护理和诊断至关重要的信息。通过针对嵌入在非结构化临床笔记中的丰富临床数据，建议的工具可以提供临床相关信息和对患者病史的上下文理解。如果成功，建议的解决方案将减轻医生的数据负担，降低丢失可能影响患者诊断或导致代价高昂的医疗错误的有价值信息的风险，并最大限度地提高对患者预后的下游影响。这个小型企业创新研究(SBIR)第一阶段项目旨在利用自然语言处理(NLP)的进步，通过自动化电子健康记录审查过程来帮助医生。潜在的创新是一个提取-抽象管道，它确定医疗记录中的哪些内容是最突出的，应该通过转换器(机器学习模型)进行总结。该项目旨在将此汇总工具推进到更具挑战性的用例中，主要汇总门诊记录，这项任务因门诊数据的大范围、临床冗余、不同的数据结构和固有的来源而变得具有挑战性，所有这些都需要在模型培训和验证中考虑在内。目标包括1)开发门诊摘要模型，并展示生成语义与参考文本高度流利匹配的摘要的能力，2)验证人工智能(AI)生成的门诊摘要的效用，以向医生提供重要价值，3)评估AI生成的摘要的能力，这些摘要通过消融研究提供与未来患者就诊相关的信息，以及4)在现有模型中纳入偏见检查。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact /commercial potential of this Small Business Innovation Research (SBIR) Phase I project is the development of a machine learning-enabled medical record summarization tool designed to provide a narrative summary that can aid physicians in patient care. On average, physicians spend just 3 minutes reviewing a patient’s medical record, and during this time they must interpret unstructured Electronic Health Records (EHR) that can make it difficult for physicians to identify information essential to patient care and diagnosis. By targeting the rich clinical data embedded in unstructured clinical notes, the proposed tool could provide clinically relevant information and a contextual understanding of a patient’s medical history. If successful, the proposed solution will reduce the data burden placed on doctors, mitigate the risk of missing valuable information that could affect patient diagnosis or lead to costly medical errors, and maximize downstream effects on patient outcomes. This Small Business Innovation Research (SBIR) Phase I project aims to leverage advances in natural language processing (NLP) to assist doctors by automating the process of electronic health record review. The underlying innovation is an extractive-abstractive pipeline that determines what content in the medical record is the most salient and should be summarized through a transformer (a machine learning model). This project aims to advance this summarization tool to more challenging use cases, primarily summarizing the outpatient record, a task made challenging by the large scope of the data, clinical redundancies, different data structures, and sources inherent to outpatient data, all of which need to be accounted for in model training and validation. Objectives include to 1) developing an outpatient summarization model and demonstrating the ability to produce summaries that semantically match reference text with a high level of fluency, 2) validating the utility of artificial intelligence (AI)-generated outpatient summaries to provide significant value to physicians, 3) evaluating the ability of AI-generated summaries that provide information relevant to future patient visit through ablation study, and 4) incorporating checks for bias in the existing model.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.