Learning alerting models for clinical care from EMR data and human knowledge

摘要 医疗差错更广泛地定义为可预防的不良临床事件。研究 表明医疗差错仍然是医疗保健和最新文献的主要挑战之一 在美国，医疗差错是主要的死亡原因之一。紧迫性和紧迫性 问题的范围促使开发旨在帮助临床医生减少 这样的错误。基于计算机的监控和警报系统，依赖于 电子病历(EMR)在这一努力中发挥着关键作用。在前几个供资周期中， 我们的团队一直在开发一种基于异常值的模型驱动的警报方法 极有可能减少医疗差错。该方法使用回溯数据来建立 从广泛的患者代表中预测医生操作的机器学习模型 各州。如果当前患者的管理操作(或其遗漏)偏离，则会发出警报 从类似患者的预期管理行动中显著受益。作为一个实际的例子 由系统生成的警报，考虑最近接受了肝脏移植的患者 并接受他克莫司作为免疫抑制剂。病人出现并发症， 接受矫正手术；然而，不经意间，他克莫司没有按照 做手术。因为没有收到预期的药物代表与预期的偏差 在类似的患者管理实践中，它是临床上的一个异常值。引发警报以重新排序 因此，药物治疗是适当的。我们目前的警报系统以静默方式部署在 在UPMC生产电子病历系统，并支持实时警报。 目前的提案将研究计划带入了一个大胆的新方向。警示模型将是 使用各种工具进行增强，包括自动评估性能和 除了包括多领域、多领域、多领域和多领域的知识库外，还包括一个适用于ICU的知识库 从电子病历派生的分辨率功能。人类专家将在确定 当实时生成警报时，警报的适当性和实用性有助于动态 知识库的增长，并评价对以下问题的解释质量 警报。最后，警报系统将在12个ICU中进行阶梯式临床试验 为了确定基于EHR的警报在向临床医生揭示时是否会修改比率和 他们行动的时机。次要终端将包括警报性能指标、流程- 相关结果和以患者为中心的结果。

Abstract Medical errors are more broadly defined as adverse clinical events that are preventable. Studies show that medical errors remain one of the key challenges of health care and recent literature ranks medical errors as one of the leading causes of death in the US. The urgency and the scope of the problem prompt the development of solutions aimed to aid clinicians in reducing such errors. Computer-based monitoring and alerting systems that rely on information in electronic medical records (EMRs) play a key role in this effort. In the previous funding cycles, our group has been developing an outlier-based model-driven alerting methodology with significant potential to reduce medical errors. The method uses retrospective data to build machine learning models that predict physician actions from a broad representation of patient states. An alert is raised if a management action (or its omission) for the current patient deviates significantly from predicted management actions for similar patients. As an example of an actual alert generated by the system, consider a patient who has recently undergone a liver transplant and receives tacrolimus as immunosuppressive agent. The patient suffers a complication and undergoes corrective surgery; however, inadvertently, tacrolimus is not reordered following the surgery. Since not receiving the expected medication represents a deviation from predicted management practice in similar patients, it is a clinical outlier. Raising an alert to reorder the medication is therefore appropriate. Our current alerting system is silently deployed on the production electronic medical record system at UPMC and supports alerting in real-time. The current proposal takes the research program in a bold new direction. Alerting models will be enhanced using a variety of tools, including automatic evaluation of performance and the inclusion of an adaptive ICU-specific knowledge-base in addition to multi-domain, multi- resolution features derived from the EMR. Human experts will play a major role in determining appropriateness and usefulness of alerts when generated in real-time, contribute to the dynamic growth of the knowledge base, and evaluate the quality of the explanations provided for the alerts. Finally, the alerting system will be deployed across 12 ICUs in a step-wedge clinical trial to determine whether EHR-based alerting, when revealed to clinicians, modifies the rate and timing of their actions. Secondary end-points will include alert performance metrics, process- related outcomes, and patient-centered outcomes.