Collaborative Research: SCH: Machine Learning Driven User Interfaces for Information Gathering and Synthesis from Medical Records

临床医生必须搜索患者过去的医疗记录，以了解患者的病情，并制定个性化的诊断和治疗计划。然而，医疗保健领域的用户界面难以操作，并且主要是来自传统临床工作流程的静态纸质表单的数字化，这种模式导致可用性差和临床医生倦怠。该项目汇集了机器学习和人机交互方面的专家，开发了一种新颖的动态上下文用户界面，将医疗笔记界面从简单的记录设备转变为可以帮助临床医生快速查找和综合信息的工具。该项目的新颖之处在于推进人类与人工智能交互的基础，并通过新颖的方法推进医疗保健领域人工智能的最新发展，这些方法可以在适当的时候为临床医生自主检索、总结和显示相关信息。该项目的影响是在国家优先考虑的卫生信息技术领域，因为它旨在实现电子健康记录的现代化。由此产生的系统将有助于防止细微的发现被忽视，患者被误诊，关键的干预措施被遗漏，最终导致发病率、死亡率和卫生保健总成本的降低。它还将减少文件负担，减轻医生的职业倦怠。以前开发患者信息上下文显示的尝试是手动的、劳动密集型的过程，依赖于领域的专业知识，既不可扩展、可维护，也不能为个人用户定制。自动化上下文显示具有挑战性，因为相关信息高度依赖于用户、患者和特定的临床上下文。由于缺乏标记的训练数据，传统的机器学习方法是不可行的。该项目开发了新的用户界面，使隐式基于使用的训练数据的大规模收集成为日常用户工作流程的一部分。具体来说，这个项目开发了一种新的信息采集界面，即“语义剪贴板”，临床医生在阅读病人过去的医疗记录和写笔记时将使用它。利用通过这个新界面收集的数据，研究人员将开发新的机器学习方法来预测应该出现在这些上下文显示中的相关信息，并根据临床场景和用户进行定制。通过这个项目和研究人员的学术教学，将培养新一代的跨学科研究人员：了解人机交互、机器学习和临床医学基本挑战的研究生，以及了解机器学习和在医疗保健中部署机器学习的微妙之处的医学研究员。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Clinicians have to search through a patient’s past medical records to contextualize the patient’s condition and reach a personalized diagnosis and treatment plan. However, user interfaces in healthcare are unwieldy to navigate and are largely a digitization of static paper forms from legacy clinical workflows, a paradigm that has contributed to poor usability and clinician burnout. This project brings together experts in machine learning and human-computer interaction to develop a novel dynamic contextual user interface that transforms the medical note-taking interface from a simple recording device to a tool that can help clinicians quickly find and synthesize information. The project’s novelties are in advancing the foundations of human-AI interaction and in advancing the state-of-the-art in artificial intelligence for health care with novel methods that autonomously retrieve, summarize, and surface relevant information for clinicians, at the right time. The project’s impacts are in an area of national priority, health IT, as it aims to modernize electronic health records. The resulting system will help prevent subtle findings from being overlooked, patients from being misdiagnosed, and critical interventions from being missed, ultimately resulting in a decrease in morbidity, mortality, and overall cost of health care. It will additionally decrease documentation burden and mitigate physician burnout. Prior attempts at developing contextual displays of patient information were manual, labor-intensive processes that relied on domain expertise and were neither scalable, maintainable, nor customized to individual users. Automating contextual displays is challenging because what information is relevant highly depends on the user, patient, and specific clinical context. Traditional machine learning approaches are infeasible because of the lack of labeled training data. This project develops new user interfaces that enable the large-scale collection of implicit usage-based training data as part of routine user workflows. Specifically, this project develops a novel information foraging interface, the ‘semantic clipboard’, which clinicians will use while reading patients’ past medical records and while writing notes. Using the data collected through this new interface, the investigators will develop new machine learning methodologies to predict the relevant pieces of information that should appear in these contextual displays, customized to the clinical scenario as well as the user. Through this project and the investigators’ academic teaching, a new generation of cross-disciplinary researchers will be educated: graduate students who understand the fundamental challenges of human computer interaction, machine learning, and clinical medicine, and medical fellows who understand machine learning and the subtleties of deploying machine learning in health care.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.