Autonomous diagnosis and management of the critically ill during air transport (ADMIT)

项目总结/摘要：呼吸不稳定（CRI）在创伤患者和其他 从创伤部位或医院中心之间转移的急性病患者。虽然 护理人员/护士（PM/RN）在使用定义的CRI抢救不稳定患者方面取得了一些成功， 协议和减少转运间严重循环休克的发生率，休克识别工具 可用的和复苏终点限于血压和心率阈值。然而，CRI是 通常未被识别，直到当患者对治疗更难治或进展到 器官损伤如果人们能够准确地预测谁，何时以及为什么这些重症患者发展CRI， 可以给予有效的先发制人的治疗，以改善护理和分诊，从而更好地利用医疗保健 资源我们已经表明，从连续非侵入性监测中获得的综合监测系统警报 所采集的监测参数与护理算法相结合，改善了降压单元（SDU）患者的治疗效果。 我们还将机器学习（ML）建模应用于我们的临床相关的出血性休克猪模型 为了表征对低血容量、出血和复苏的反应，预测哪些动物会或 在低血容量期间不会崩溃，并且比传统的方法提前5分钟发现隐匿性出血 监测.我们现在建议将我们的工作应用于脆弱的STAT MedEvac空运患者。我们将 在我们现有的> 5，000名患者的STAT MedEvac数据库中验证这些方法， 连续无创监测波形的空运危重病人连接到他们的主要 电子健康记录（EHR）。这种级别的患者信息和粒度与 治疗数据和患者结果是前所未有的。我们将扩展我们的分析，以包括更复杂的 CRI、更丰富的数据、更深入的分析以及更大的重症患者库，将我们的 经过验证的功能性血液动力学监测（FHM）原理，用于病理生理诊断和复苏 通过非侵入性监测来实施个性化复苏。我们将同时运行两个 具体目标。首先，我们将通过卡内基梅隆大学Auton实验室开发多变量模型 通过ML数据驱动的分类技术来预测CRI。我们将首先在现有的猪上进行此操作 出血性休克模型数据（n=60），然后在我们的STAT MedEvac数据集上与EHR（n > 5，000）相关联 患者），确定所需的最低数据（措施，采样频率，观察持续时间）， 稳健地识别与健康状况的偏差、可能的CRI原因和对治疗的反应（复苏终点）， 以及附加变量、分析、交付周期和采样频率的增量效益， CRI和对治疗的反应，并检查模型简约性和特异性之间的权衡。第二、 我们将评估我们现有的临床决策支持（CDS）工具，以与FHM原则和ML接口， 定义的相互作用，并首先在我们的猪出血性休克复苏中进行计算机模拟试验，然后在我们的 STAT MedEvac数据，随后在年度期间对飞行机组PM/RN（n=160）进行前瞻性人体模拟 协议和受益培训，根据诊断准确性定义有效性，诊断时间， 干预选择准确性和干预时间。这个迭代过程将修改现有的CDS 平台转换为更适合航空运输场景的平台。最后，我们将评估结果 半自主管理方案最初在100名STAT MedEvac患者和10名 急诊科创伤患者，然后在最终100个STAT MedEvac中前瞻性地通过主动CDS 患者我们将前瞻性地分析这些校准CDS工具的有效性，以预测 各种ML模型，并将最佳、最实用和最简约的预测模型应用于临床护理 根据患者人数、病理过程和支持人员进行运输。

Project Summary/Abstract: Cardiorespiratory instability (CRI) is common in trauma patients and other acutely ill patients being transferred from trauma sites or between hospital centers. Although paramedics/nurses (PM/RN) have some success in rescuing unstable patients with CRI using defined protocols and decrease incidence of inter-transport severe circulatory shock, the shock recognition tools available and resuscitation endpoints are limited to blood pressure and heart rate thresholds. However, CRI is often unrecognized until it is well established when patients are more refractory to treatment, or progressed to organ injury. If one could accurately predict who, when and why these critically ill patients develop CRI, then effective preemptive treatments could be given to improve care and triage resulting in better use of healthcare resources. We have shown that an integrated monitoring system alert obtained from continuous noninvasively acquired monitoring parameters coupled to a care algorithm improved step-down unit (SDU) patient outcomes. We also applied machine learning (ML) modeling to our clinically-relevant porcine model of hemorrhagic shock to characterize responses to hypovolemia, hemorrhage, and resuscitation, predict which animals would or would not collapse during hypovolemia, and identify occult bleeding 5 minutes earlier than with traditional monitoring. We now propose to apply our work to vulnerable STAT MedEvac air transported patients. We will validate these approaches in our existing >5,000 patient STAT MedEvac database, containing highly granular continuous non-invasive monitoring waveforms of air transported critically ill patients linked to their primary care and inpatient electronic health records (EHR). This level of patient information and granularity linked to treatment data and patient outcomes is unprecedented. We will extend our analysis to include more complex CRI, richer data, deeper analytics, and larger libraries of critically ill patients while in air transport, linking our proven Functional Hemodynamic Monitoring (FHM) principles for pathophysiologic diagnosis and resuscitation with non-invasive monitoring to operationalize personalized resuscitation. We will concurrently running two specific aims. First, we will develop through the Carnegie Melon University Auton Lab multivariable models through ML data-driven classification techniques to predict CRI. We will do this initially on our existing porcine hemorrhagic shock model data (n=60) and then on our STAT MedEvac dataset linked to EHR (n >5,000 patients), determining the minimal data (measures, sampling frequency, observation duration) required to robustly identify deviation from health, likely CRI cause, and response to treatment (endpoint of resuscitation), as well as the incremental benefit of additional variables, analysis, lead-time and sampling frequency to predict CRI and response to treatment, and examine the trade-offs between model parsimony and specificity. Second, we will evaluate our existing clinical decision support (CDS) tools to interface with FHM principles and ML- defined interactions, and trial this in silico first on our porcine hemorrhagic shock resuscitation, then on our STAT MedEvac data, followed by prospective human simulation on flight crew PM/RN (n=160) during annual training for agreement and benefit, defining effectiveness based on diagnosis accuracy, time to diagnosis, intervention choice accuracy and time to intervention. This iterative process will modify the existing CDS platform into one more specifically suited for air transport scenarios. Finally, we will evaluate the resultant semi-autonomous management protocol initially in retrospect in 100 STAT MedEvac patients and 10 Emergency Department trauma patients and then prospectively by active CDS in a final 100 STAT MedEvac patients. We will prospectively analyze the effectiveness of these calibrated CDS tools for predictive ability of the various ML models and apply the best, most practical and parsimonious predictive models for clinical care during transport based on patient population, pathological processes and support staff.