A Modeling and Control Framework for Early Detection of Adverse Clinical States

重症监护病房是所有医院中死亡率最高的病房。这些重症患者同时接受多种复杂的干预措施，而且护理非常复杂，他们极易受到医疗差错和不良后果的影响。重症监护病人是医院中使用仪器最多的病人。生理信号是用许多不同类型的传感器收集的。这些传感器信号反映了患者潜在的动态、综合的生理状态，因此是高度复杂和相互关联的。重症监护医生面临的最大挑战是，这些数据的数量和复杂性推动了他们认知吸收的极限，并将其与患者的整体生理状态联系起来。他们每时每刻都面对着大数据，但他们必须快速解读并采取行动。他们缺乏这样做的工具。在这个项目中，目标是开发和应用一类新的算法，称为最佳变化点检测算法，来检测患者状态何时从一种临床状态变化到另一种临床状态。复杂算法在医疗保健中的应用不仅会改变重症监护病房的治疗，而且还会被带到课堂上，让选修计算医学的本科生学习。研究小组计划开发自动化计算方法，用于处理来自重症监护患者传感器的生理时间序列数据，以快速检测临床状态的变化。这些方法包括：(i)改进了患者状态特征的选择；（ii）基于这些特征检测患者状态转变的最优控制算法。该方法的一个创新之处在于将状态转移检测问题重新表述为控制理论和贝叶斯最优序列决策领域的最优变点问题。这反过来又使我们能够通过首先定义反映性能目标的成本函数（例如，最大化灵敏度，最小化误报），然后开发最小化该成本的检测规则来推导出最佳检测器。研究人员已经证明，与其他最先进的方法相比，这种方法可以将检测癫痫发作的时间缩短50%。另一个创新在于如何从生理波形中计算特征。拟议研究的目标是确定最佳变化点算法与这些新特征相结合，应用于多种类型的生理时间序列数据的分析，是否可以支持早期检测重症监护病房患者临床状况的变化。将开发一门新的课程或课程模块，介绍统计和机制模型估计以及简单的早期检测算法。本课程将使生物医学工程专业的学生接触到具有临床相关性的生物系统的实验数据、基于生物物理的模型和统计模型。

Critical care units are the highest mortality units in any hospital. These severely ill patients undergo multiple complex interventions at the same time, and care is so complex that they are extremely vulnerable to medical errors and adverse outcomes. Critical care patients are the most heavily instrumented patients in the hospital. Physiological signals are collected using many different types of sensors. These sensor signals reflect the underlying dynamic, integrated physiological state of the patient and are thus highly complex and inter-related. The biggest challenge faced by critical care physicians is that the amount and complexity of these data push the limits of what they can cognitively assimilate and relate to the overall physiological status of their patients. They are confronted by Big Data at every moment, yet they must interpret and act on them quickly. They lack the tools to do this. In this project, a goal is to develop and apply a novel class of algorithms, known as optimal change-point detection algorithms, to the problem of detecting when a patients state changes from one clinical condition to another. The application of sophisticated algorithms to healthcare will not only transform treatment in critical care units, but will be brought to the classroom to undergraduate students minoring in Computational Medicine.The research team plans to develop automated computational methods for processing physiological time series data from critical care patient sensors to quickly detect changes in clinical state. These methods will include; (i) improved selection of features that characterize patient state; (ii) optimal control algorithms to detect transitions of patient state based on these features. One innovation in the proposed approach is a re-formulated the state transition detection problem as an optimal change-point problem from the fields of controls-theory and Bayesian optimal sequential decision making. This in turn has enabled us to derive an optimal detector by first defining a cost function that reflects performance goals (e.g. maximize sensitivity, minimize false positives), and then developing the detection rule that minimizes this cost. The researchers have demonstrated that this approach decreases time to detection of epileptic seizure onset by 50% relative to other state of the art methods. Another innovation lies in how to compute features from physiological waveforms. The goal of proposed research is to determine if optimal-change point algorithms in conjunction with these new features applied to the analysis of multiple types of physiological time series data can support earlier detection of changes in the clinical conditions of patients in the critical care unit. A new course or course module that introduces statistical and mechanistic model estimation and simple early detection algorithms will be developed. This course will expose biomedical engineering students to experimental data, biophysical-based models, and statistical models of biological systems that have clinical relevance.