Timely Response to In-Hospital Deterioration Through Design of Actionable Augmented Intelligence

摘要 用于预测临床结果的增强智能(AI)模型的发展呈指数级增长。 自动化临床监测，以协助及早发现院内恶化情况，如脓毒症和急性 肾损伤是一种很有前途的人工智能应用。每年有多达30万名美国医院患者死于 脓毒症、急性呼吸道感染等问题以及5%或更多的死亡是可以预防的。更多的患者遭受 伤害或额外成本，作为延迟响应的后遗症。与传统的基于规则的风险相比 预测，使用机器学习等方法的高级人工智能模型展示了更高的可靠性 预测脓毒症和AKI。这些系统在实践中的有效性很可能取决于人工智能如何风险 信息被集成到临床工作流程和技术中，但我们还没有意识到要设计或 评估有效的医院内人工智能风险信息呈现和用户交互。众所周知， 可解释的人工智能是可取的，但需要解释什么以及如何有效和高效地进行解释则不可取 为人所知。有必要了解终端用户对人工智能在特定临床环境中的价值的看法。 我们将借鉴我们团队最近在有效临床展示设计、人类人工智能理论模型方面的研究 表现，人工智能设计原则的应用，以及以人为中心的设计方法在 设计和评估有效的方法，以支持对脓毒症和AKI风险的及时反应。我们的初选 目标是：确定影响临床医生感知人工智能有用性的因素，生成设计 有效的健康风险监测人-人工智能交互的原则，并设计人-人工智能用户界面 在应对败血症和AKI时，有意义地提高人类人工智能的表现。在目标1中，我们将制定一个 时间推理AI模型预测脓毒症和AKI的院内发展。我们将应用此模式 回顾患者数据，作为研究活动的背景。使用图表回顾，我们将量化 人类-人工智能系统性能的现实指标，考虑人工智能模型是否 在临床团队怀疑事件或对事件采取行动之前预测恶化。在目标2中，我们将 在访问临床医生的同时回顾患者在住院期间病情变化的时间进程， 包括人工智能生成的风险信息。我们将收集关于影响临床医生的因素的定性数据 对于早期识别患者问题的人工智能信息的有用性的看法。在《目标3》中， 我们将与临床医生开展参与式设计活动，设计有效的以人为中心的人工智能显示器和 相互作用以支持院内败血症和AKI的早期反应。最后，在目标4中，使用模拟真实感 患者护理任务，并与传统的患者信息技术相比，我们将评估 以人为中心的人工智能设计--人工智能性能。我们将生成人工智能交互设计 健康风险监测指导意见。我们的发现有望为人类-人工智能交互设计带来创新 电子健康记录(EHR)以及保健监测和通信技术。

Abstract Development of augmented intelligence (AI) models for predicting clinical outcomes is growing exponentially. Automated clinical surveillance to assist in early detection of in-hospital deterioration such as sepsis and acute kidney injury (AKI) is a promising AI application. As many as 300,000 US hospital patients die each year from problems like sepsis and AKI and 5% or more of these deaths are preventable. Many more patients suffer harm or additional costs as a sequelae to delayed response. Compared to traditional rule-based risk predictions, advanced AI models using methods such as machine learning demonstrate improved reliability of predicting sepsis and AKI. The effectiveness of these systems in practice will likely depend on how AI risk information is integrated into clinical workflow and technologies, yet we are not aware of research to design or evaluate effective in-hospital AI risk information presentation and user interaction. It is widely known that explainable AI is desirable, but what needs to be explained and how to do it effectively and efficiently is not known. There is a need to understand end user perspectives on the value of AI for specific clinical contexts. We will draw on our team's recent research on effective clinical display design, theoretical models of human-AI performance, application of human-AI design principles, and application of human-centered design methods to design and evaluate effective approaches to support timely response to sepsis and AKI risk. Our primary objectives are to: identify factors that influence clinicians' perceptions of AI usefulness, generate design principles for effective health risk surveillance human-AI interaction, and design human-AI user interfaces that meaningfully improve human-AI performance when responding to sepsis and AKI. In Aim 1, we will develop a temporal reasoning AI model for predicting in-hospital development of sepsis and AKI. We will apply this model to retrospective patient data to serve as context for research activities. Using chart review, we will quantify realistic metrics of human-AI system performance that take into account whether the AI model would have predicted deterioration before the clinical team suspected or acted in response to the event. In Aim 2, we will interview clinicians while reviewing temporal progression of a patient's change in condition over their stay, including AI generated risk information. We will gather qualitative data on factors that influence clinicians' perceptions of usefulness of AI information toward the goal of early identification of patient problems. In Aim 3, we will conduct participatory design activities with clinicians to design effective human-centered AI display and interaction to support early response to in-hospital sepsis and AKI. Finally, in Aim 4, using simulated realistic patient care tasks and comparing to traditional patient information technologies, we will evaluate the impact of human-centered AI designs on human-AI performance. We will generate human-AI interaction design guidance for health risk surveillance. Our findings are expected to innovate design for human-AI interaction in electronic health records (EHRs) and health-care monitoring and communication technologies.