GeneRAlizable Sepsis Phenotyping (GRASP) using Electronic Health Records and Continuous Monitoring Sensors

摘要 脓毒症，一种异质综合征，其特征是全身炎症，由身体的 对感染的反应是医院治疗的最昂贵和最致命的疾病，超过27万人 仅在美国就有与脓毒症相关的死亡病例。未经治疗的脓毒症可能导致血液扩张和渗漏 血管和需要血管活性药物的严重低血压(也称为感染性休克)，以及最终对 肾、肺和肝脏(又称器官损伤)死亡率超过40%。成功的预防和 脓毒症、感染性休克和器官损伤的治疗依赖于临床医生预测和治疗的能力 评估风险，并实施正确的救命治疗(如抗生素、液体和血管加压剂) 在正确的时间。近年来，数据驱动建模已被证明能够对脓毒症进行早期预测 以及揭示脓毒症的簇状(或表型)，这可能有助于个性化治疗 干预措施。然而，跨越临床研究和改善患者之间的翻译鸿沟 护理还需要通过更好的数据整合，在不同的护理水平上解决1)“数据沙漠”问题， 更智能的实验室订购和持续监控可穿戴传感器的利用；2)互操作性和 临床数据和分析的可移植性；3)原则性传播和实施研究；4) 对下一代照料者进行有效利用先进分析工具的教育。 拟议的研究计划建立在Pi的K01早期职业发展奖的基础上，重点是 脓毒症预测分析算法(包括每小时EHR数据)的多中心开发和验证 横跨五个地区超过500,000名住院患者的急诊和住院会诊 医疗保健系统)。借鉴领域适应和多任务学习的最新进展 (机器学习的子领域)，该项目旨在发现可概括的动态表型，这些表型是 与脓毒症、感染性休克和下游器官损伤的预测和处理直接相关。 我们建议使用来自床边设备的高分辨率数据来增强基于EHR的分析(例如， 监护仪、呼吸机、透析和静脉输液泵)和可穿戴设备(例如，持续血压和乳酸 传感器)，以弥补监测方面的现有差距。此外，该计划旨在推进FHIR(快速 医疗保健互操作性资源)和OMOP(观察性医疗结果伙伴关系) 通过实施针对高分辨率数据源的特定资源来实现互操作性标准。 最后，这项研究计划将与我们的传播和 实施和医院质量改进团队以确保及早评估可用性、障碍和 实施和有效的教育，以最大限度地发挥潜在的临床影响。

Abstract Sepsis, a heterogeneous syndrome characterized by whole-body inflammation caused by the body's response to an infection, is the most expensive and deadly condition treated in hospitals, with over 270,000 cases of sepsis-related deaths in the U.S. alone. Untreated sepsis may result in dilated and leaky blood vessels and severe hypotension requiring vasoactive medications (aka septic shock), and eventual injury to kidneys, lungs, and liver (aka organ injury) with mortality rates in excess of 40%. Successful prevention and management of sepsis, septic shock, and organ injury rely on the ability of clinicians to anticipate and estimate the risk, and administer the right life-saving treatments (e.g., antibiotics, fluids and vasopressors) at the right time. In recent years, data-driven modeling has been shown to enable early prediction of sepsis and to reveal clusters (or phenotypes) of sepsis, which may help with personalizing therapeutic interventions. However, crossing the translational chasm between clinical research and improving patient care also requires addressing 1) `data deserts' at different levels of care through better data integration, smarter lab ordering, and utilization of continuous monitoring wearable sensors; 2) interoperability and portability of clinical data and analytics; 3) principled dissemination and implementation studies; and 4) education of the next generation of caregivers to effectively utilize advanced analytical tools. The proposed research program builds upon PI's K01 early career development award focused on multicenter development and validation of sepsis predictive analytic algorithms (including hourly EHR data spanning ED and inpatient encounters from over 500,000 hospitalized patients across five district healthcare systems). Drawing insights from recent advances in domain adaptation and multi-task learning (sub-fields of machine learning), this project aims to discover generalizable dynamic phenotypes that are directly relevant to the prediction and management of sepsis, septic shock, and downstream organ injury. We propose to augment EHR-based analytics with high-resolution data from bedside devices (e.g., monitors, ventilators, dialysis, and IV pumps) and wearables (e.g., continuous blood pressure and lactate sensors) to address existing gaps in monitoring. Additionally, this program aims at advancing FHIR (Fast Healthcare Interoperability Resources) and OMOP (Observational Medical Outcomes Partnership) interoperability standards through the implementation of specific resources for high-resolution data sources. Finally, this research program will be conducted in close collaboration with our dissemination and implementation and hospital quality improvement teams to ensure early assessment of usability, barriers to implementation, and effective education to maximize the potential for clinical impact.