Solving Sepsis: Early Identification and Prompt Management Using Machine Learning

摘要 这一快速跟踪STTR应用程序计划增强、验证和扩展Sepsis Watch，这是一项深入的 利用急诊数据构建学习型脓毒症检测和管理系统 杜克大学医院(DUH)。该提案将扩展和增强 对急诊室、普通住院病房和重症监护病房(ICU)设置的脓毒症观察 美国的多个医疗系统。早期诊断和及时治疗 脓毒症可以提高死亡率和发病率，但早期发现仍然难以实现。脓毒症 DUH ED中的手表集成将3小时脓毒症捆绑包的遵从性提高了12% 而6小时的脓毒症束则增加了18%。该系统将严重脓毒症的死亡率降低了15% 感染性休克死亡率下降22%。这项提议旨在将Sepsis Watch转变为 可扩展的解决方案，可在其他卫生系统和急诊科以外的环境中复制此类结果。 在第一阶段，我们建议通过回顾分析来自两个国家的数据进行外部验证 独立的医疗系统。第1阶段将允许我们自动执行数据质量检查和接收 来自不同医疗系统的大规模流程，因为我们从至少200,000， 在两年的时间里相遇。我们将向临床医生提供模型预测， 医院分析将脓毒症观察整合到临床护理中的潜在影响。在第二阶段， 我们建议从第一阶段开始在每一家医院进行临时验证。这将使我们能够 将数据记录实时接收到Sepsis Watch的设计方式与 电子健康记录(EHR)供应商系统。我们将优化机器学习模型 使用阶段1调查结果提高每个位置的性能，同时评估联合和 将来自不同医院的数据合并在一起的集中学习方法。模型 还将评估使用不同输入集合的变化，并将构建模型以 三个金标准脓毒症定义，包括脓毒症-3、CMS SEP-1脓毒症和CDC成人 败血症事件。在为期6个月的时间验证期间，我们还将推广脓毒症观察 用户界面和工作流程，在运行过程中向每家医院的临床医生寻求反馈 静默模式。这将允许Sepsis Watch可配置为各种临床工作流程。 第二阶段的优化模型和用户界面应该允许Sepsis Watch无缝 整合到每个医院的常规临床护理中，然后整合到每个医院的其他医院中 这两个医疗系统，并最终到美国的任何医疗系统。

Abstract This fast-track STTR application proposes to enhance, validate, and scale Sepsis Watch, a deep learning sepsis detection and management system built using data from the Emergency Department (ED) Duke University Hospital (DUH). The proposal will extend and enhance Sepsis Watch to EDs, general inpatient wards, and intensive care unit (ICU) settings across multiple health systems in the United States. While early diagnosis and prompt treatment of sepsis can improve mortality and morbidity, early detection has remained elusive. The Sepsis Watch integration in the DUH ED improved compliance with the 3-hour sepsis bundle by 12% and the 6-hour sepsis bundle by 18%. The system reduced mortality for severe sepsis by 15% and mortality for septic shock by 22%. This proposal seeks to transform Sepsis Watch into a scalable solution to replicate such results at other health systems and in settings beyond the ED. In Phase I, we propose external validation through a retrospective analysis of data from two separate health systems. Phase 1 will let us automate data quality checks and ingestion processes at scale from different health systems as we curate data from at least 200,000 encounters over a 2-year period. We will present model predictions to clinicians from each hospital to analyze potential impact of integrating Sepsis Watch into clinical care. In Phase II, we propose conducting temporal validation at each hospital from Phase I. This will allow us to design real-time ingestion of data records into Sepsis Watch in a manner that is agnostic to electronic health record (EHR) vendor systems. We will optimize the machine learning model using Phase 1 findings to improve performance at each location while assessing federated and centralized learning approaches that incorporate data from different hospitals. Models variations that utilize different sets of inputs will also be assessed and models will be built to three gold-standard sepsis definitions, including Sepsis-3, CMS SEP-1 sepsis, and CDC Adult Sepsis Event. During the 6-month temporal validation we will also generalize the Sepsis Watch user-interface and workflow by seeking feedback from clinicians at each hospital as it is run in silent mode. This will allow Sepsis Watch to be configurable to various clinical workflows. The optimized model and user-interface in Phase 2 should allow Sepsis Watch to be seamlessly integrated into routine clinical care in each hospital and then into other hospitals within each of the two health systems and eventually to any health system in the US.