Sepsis Physiomarkers for Appropriate Risk Knowledge of monitored patients in the ICU (SPARK-ICU)

项目摘要 重症患者进入ICU谁发展继发感染和败血症，可面临高达五倍， 与非败血症患者相比，死亡风险增加。大多数发展为 二次感染在入院时病情更严重，因此需要更多的资源。 用于预测脓毒症的传统机器学习算法在很大程度上集中并依赖于使用 来自电子病历（EMR）的结构化数据，但电子病历主要是作为账单开发的 机制和临床工作流程的审计日志。因此，数据的大部分结构和可用性往往是 时间延迟，容易出现手动输入错误，各种机构，个人和培训偏见的偏见， 并且最终包含大量的缺失数据。在这个提议中，我们试图发现新的 从连续生理数据流中提取的“生理标记”，从非人类衍生数据中生成 预测败血症在这一关键人群中的发病。使用这种常规收集的数据，沿着 从EMR中提取的常见临床指标，我们提出生成鲁棒的机器学习算法 这可以更普遍、可重复，并消除手动数据输入的偏见和陷阱。我们 这些模型不仅可以提醒临床医生急性和重症患者有风险， 实时发展脓毒症，但也研究干预的有效性，如容量反应性 并支持败血症新亚型的发现。其次，现有的许多关于预测的文献 脓毒症的模型集中在普通病房的住院患者，然而，预测脓毒症发作的模型 在进入ICU后发生继发感染的患者中，脓毒症的发生率有限。在我们以前 工作，我们已经证明，从连续的数字数据流中发现的标记可以更早地通知 预测儿童和成人的脓毒症。然而，这些分析并没有使用高保真数据， 波形，其中封装了丰富的生理特征。因此，通过强调发现 这种新的标记，并通过应用数据驱动的学习算法，我们希望开发 算法和工具，以提高我们的理解不断变化的生理动力学脓毒症， 生病的病人。在这个拟议的计划中，我们将整合一些独特的专业知识， 跨越信号处理、数学、计算机科学和医学，开发复杂的工具， 分析这些数据以揭示有意义的见解。简而言之，我们将贡献有关角色的重要知识， 以及复杂生理相互作用的实用性，这些生理相互作用目前在临床实践中大量存在， 很少用于临床决策。

Project Summary Critically ill patients admitted to the ICU who develop secondary infection and sepsis, can face up to a five-fold increase in the risk for death when compared to non-sepsis patients. The majority of patients who developed secondary infections are more critically ill at admission and therefore require significantly greater resources. Traditional machine learning algorithms for predicting sepsis has been largely focused and relied on the use of structured data from the electronic medical record (EMR), however the EMR was developed largely as a billing mechanism and an audit log for clinical workflow. Hence, much of the structure and availability of data are often time-delayed, prone to errors from manual entry, biases from various institutional, personal and training biases, and finally contain a significant amount of missing data. In this proposal, we seek to discover novel `physiomarkers' extracted from continuous physiological data streams, generated from non-human derived data sources, that predict the onset of sepsis in this critical population. Using such routinely collected data, along with common clinical indicators extracted from the EMR, we propose to generate robust machine learning algorithms that can be more generalized, reproducible and removed from the biases and pitfalls of manual data entry. We propose that such classes of models not only may alert clinicians to acute and critically ill patients at risk for developing sepsis in real-time, but also investigate intervention effectiveness, such as volume responsiveness and support the discovery of novel sub-types of sepsis. Secondly, much of the existing literature on predictive models for sepsis focus on hospitalized patients in the general ward, however, models that predict the onset of sepsis among patients who developed secondary infections after admission to the ICU is limited. In our previous work, we have demonstrated that markers discovered from continuous numeric data streams can inform earlier prediction of sepsis in children and adults. However, those analysis did not use high-fidelity data from the waveforms, which encapsulate rich characteristics of physiology. Therefore, by emphasizing the discovery of such novel markers and through the application of data-driven learning algorithms, we expect to develop algorithms and tools that improve our understanding of the changing physiologic dynamics of sepsis in critically ill patient. In this proposed program, we will integrate knowledge across a number of distinctive expertise that spans signal processing, mathematics, computer science and medicine to develop sophisticated tools that can analyze such data to reveal meaningful insight. In short, we will contribute significant knowledge about the role and utility of complex physiological interactions that are at present abundantly available in clinical practice but seldom used for clinical decision making.