Predicting Patient Instability Noninvasively for Nursing Care-Two (PPINNC-2)

 描述（由申请方提供）：降压单元（SDU）中的患者接受连续无创生命体征（VS）监测，以帮助护士检测需要诊断和/或治疗反应的心肺不稳定（CRI）。然而，我们的数据和其他数据显示，护士并不容易识别CRI，也不一定及时做出反应。此外，75%的监测SDU患者从未变得不稳定，分散监测远离那些需要的人。目前的护理监督策略是不精确和无针对性的，导致无法检测CRI，并未能抢救需要干预的患者。智能临床决策支持系统（SDSS）可以持续处理来自不同来源的复杂数据，可以应用检测算法提醒临床医生正在发生的事件（即时预报）和未来发生的事件（预测），以重新将临床医生的重点放在高风险患者上，并更早地应用支持性护理，甚至预防CRI。在我们之前的R 01中，我们使用机器学习（ML）开发了一个原型SDSS来处理时间序列数据，并系统地学习近期或远期事件之前的数据模式。在这项研究中，我们：1）从我们现有的1/20 Hz生理监测数据中组装了一个复杂的专家注释多维数据集作为学习平台; 2）开发并验证了用于动态CRI预测的ML模型，测试了在CRI即时预报中最有效的模型; 3）使用SDSS为护士开发了早期原型CRI预测图形用户界面（GUI）。重要的是，我们的模型还能够以高灵敏度和特异性实时区分真实的CRI和伪影。我们现在建议在这项工作的基础上，通过以下方式改进我们的模型：1）前瞻性地收集多维高分辨率来自两个不同医疗中心SDU的非侵入性生理监测和电子健康记录（EHR）数据的（100- 250 Hz）数据集，以创建增强的和可推广的CRI即时预报能力，2）使用我们现有的和正在扩展的多中心临床数据集来精确预测CRI事件和在线伪影辨别，以及3）创建用于实时精确CRI临近预报和预报警报的鲁棒SDSS平台;在图形用户界面（GUI）中呈现警报，在模拟设置中迭代地开发GUI;然后在SDU中试验其使用。允许护士走向SDSS支持的精确护理监测，并知道谁会变得不稳定，当他们这样做，以及为什么-所有在公开的不稳定表现之前-可以将CRI护理从反应性转变为先发制人。开发一种灵敏、特异、简约和临床实用的方法来预测患者不稳定性，对于降低可预防的发病率和死亡率、消除警报疲劳、改善患者安全、护理（监测频率、病例负荷和混合、工作人员分配）和护理提供系统（分诊、床位分配、预防不良事件）具有重要意义。

 DESCRIPTION (provided by applicant): Patients in step-down units (SDU) undergo continuous noninvasive vital sign (VS) monitoring to facilitate nurse detection of cardiorespiratory instability (CRI) in need of a diagnostic and/or therapeutic response. Yet, our data and others show nurses do not readily identify CRI nor necessarily react in a timely fashion. Further, 75% of monitored SDU patients never become unstable, diffusing surveillance away from those in need. Current nursing surveillance strategies are imprecise and untargeted, leading to failure to detect CRI, and failure to rescue patients needing intervention. Smart clinical decision support systems (SDSS) that continuously process complex data from disparate sources can apply detection algorithms to alert clinicians of ongoing events (nowcasting) and those developing in future (forecasting) to refocus clinicians on high-risk patients, and apply supportive care earlier, or even care to prevent the CRI. In our prior R01 we developed a prototype SDSS using machine learning (ML) to process time series data and learn data patterns systematically preceding events in the near or far term. In that study, we: 1) assembled a complex expert-annotated multidimensional dataset from our existing 1/20 Hz physiologic monitoring data as the learning platform; 2) developed and validated ML models for dynamic CRI prediction, tested which were most effective in CRI nowcasting; and 3) used the SDSS to develop an early prototype CRI prediction graphical user interface (GUI) for nurses. Importantly, our models were also able to discriminate between real CRI and artifact in real-time with high sensitivity and specificity. We now propose to build on this work and enhance our models by: 1) prospectively collecting a multidimensional high-resolution (100-250Hz) dataset of non-invasive physiologic monitoring and electronic health record (EHR) data from two different medical center SDUs to create enhanced and generalizable CRI nowcasting capability, 2) using our existing and the expanding multicenter clinical datasets to precisely forecast CRI events and online artifact discrimination, and 3) create a robust SDSS platform for real-time precision CRI nowcast and forecast alerting; present alerts in a graphical user interface (GUI), iteratively develop the GUI in the simulation setting; and then pilot its use in a SDU. Permitting nurses to move toward SDSS-supported precision nursing surveillance, and know who will become unstable, when they will do so, and why-all in advance of overt instability manifestation-can shift CRI nursing care from reactive to preemptive. Developing a sensitive, specific, parsimonious and clinically practical means to predict patient instability has important implications for reducing preventable morbidity and mortality, eliminating alarm fatigue, improving patient safety, nursing care (monitoring frequency, case load and mixture, staff allocation) and care delivery systems (triage, bed allocation, prevention of adverse events).