Characterizing patients at risk for sepsis through Big Data

通过大数据确定脓毒症风险患者的特征 摘要 KL2研究提案的目标是创建现有数据驱动的脓毒症算法的扩展，即 人工智能败血症专家(AISE)，通过预测脓毒症特定器官衰竭的类型和顺序 使用电子病历中的临床数据，然后确定 添加从心血管波形(动脉波形和 心电波)和小分子代谢物随时间变化的数据提供了一些机制 背景。来自AISE的实时使用的最重要的数据(特征)将被用作模糊k- 使用回溯收集的数据的均值聚类算法(从脓毒症中拯救器官，或SOS)， 旨在更好地描述因器官衰竭而有脓毒症风险的患者。我已经选择了三个器官 具体：休克、急性呼吸衰竭和急性肾损伤(AKI)。主成分分析(PCA) 来自2375名患有脓毒症的ICU患者的数据将被投影到患者的新视觉表示上 根据不同类型器官衰竭的风险(轨迹)进行表型分析。我会确认这个SOS是否 基于SOS的算法可以在12小时内准确预测新的器官衰竭。 为了更好地了解特定功能对器官衰竭的影响，我将测试每一种高- 分辨率特征和代谢组学数据预测感染性休克。在那些出现感染性休克的人中，我 将测量从ICU开始到休克发作的所有九个高分辨率特征 将这些变化与那些发生败血症但不发生感染性休克的人和那些没有发生败血症的人进行比较。 败血症。然后，我会看看高分辨率数据的集体添加是否会改善感染性休克的表现 预测。最后，我将进行一项前瞻性的观察性研究，以收集60- 患者研究确定在SOS中加入代谢组学数据预测脓毒症的渐进性改善 电击，在仅有EMR和波形数据的SOS上。这项工作的结果将为以下工作提供初步数据 进一步的职业发展和国家卫生研究院的资助。长期目标将是建立一种优化 适当的治疗时机，从而减少脓毒症和相关器官衰竭的发生率。 作为一名K23候选人，我将利用这个奖项来获得正式的教学培训和更多的实践经验 机器学习、信号处理、代谢组学分析。我将寻求有针对性的培训，以补充 我作为临床试验者的经验，这样我就可以设计高质量的研究，为 重症监护研究和实践。我的首要职业目标是成为大型应用程序的领导者 对危重病人的数据分析以预测疾病的进展，特别是脓毒症。埃默里 环境是发展这些能力的理想场所。Emory Healthcare容纳了200多名医疗人员， 外科和专科ICU床位，其中许多床位都“连接”以存储流数据。除 身体资源，我的发展将得到由Greg Martin博士领导的卓越的指导团队的推动。

Characterizing Patients at Risk for Sepsis Through Big Data SUMMARY The goal of this KL2 research proposal is create an extension of an existing data-driven sepsis algorithm, the artificial intelligence sepsis expert (AISE), by forecasting the type and sequence of sepsis-specific organ failure using clinical data from the electronic medical record, then identifying the incremental benefit received by adding high-resolution data derived from cardiovascular waveforms (arterial waveform and electrocardiographic waves), and small molecule metabolite data over time to provide some mechanistic context. The most important data (features) used in real-time from AISE will be used as inputs for a fuzzy k- means clustering algorithm (“Saving Organs from Sepsis”, or SOS) using retrospectively-collected data, designed to better characterize patients at risk for sepsis by their organ failure. I have selected three organs in particular: shock, acute respiratory failure, and acute kidney injury (AKI). Principal component analyses (PCA) data from 2,375 ICU patients with sepsis will be projected onto a novel visual representation for patient phenotyping based on risk of (trajectory toward) different types of organ failure. I will identify if this SOS algorithm can accurately forecast new organ failure within 12 hours based on SOS. To better understand the impact of specific features on organ failure, I will test the ability of each high- resolution features, and metabolomics data to forecast septic shock. Among those who develop septic shock, I will measure all nine high-resolution features from the beginning of the ICU stay up to shock onset and compare those changes to those who develop sepsis but not septic shock, and those who do not develop sepsis. I will then see if the collective addition of high-resolution data improves performance of septic shock forecasting. Finally, I will conduct a prospective observational study to collect metabolomic information in a 60- patient study to identify the incremental improvement of adding metabolomics data to SOS for predicting septic shock, over SOS with just EMR and waveform data. The results of this work will provide preliminary data for further career development and NIH-funding. The long-term goal would be to build a model that optimizes the timing of appropriate therapy, thus decreasing the incidence of sepsis and associated organ failure. As a K23 candidate, I will use this award to acquire formal didactic training and more hands-on experience in machine learning, signal processing, metabolomics analysis. I will seek focused training that will complement my experience as a clinical trialist so that I can design high-quality studies to contribute to Big Data analytics in critical care research and practice. My overarching career goal is to become a leader in the application of Big Data analysis of critically ill patients to predict progression of disease, specifically sepsis. The Emory environment is an ideal place to develop these capabilities. Emory Healthcare houses over 200 medical, surgical, and subspecialty ICU beds, many of which are “wired” to store streaming data. In addition to the physical resources, my development will be enhanced by my superb mentorship team, led by Dr. Greg Martin.