Prediction of Clinical Deterioration Using a Bayesian Belief System

项目摘要/摘要 大约5%-10%的住院患者在入院后临床病情显著恶化，导致 要么转到重症监护病房(ICU)，要么发生“代码”事件(即心脏或肺骤停)。延迟 查明这些事件会导致发病率和死亡率的增加。不幸的是，现有的预测 模型生成的每个真阳性警报都会导致多个错误警报。除了每一项 在过去的一年里，新的监测系统被引入，产生更多的错误警报，导致警报 疲劳一直与病人的死亡有关。 这个有指导的职业发展提案的目标是开发和评估新的计算能力 算法可以更早和更准确地预测住院患者的临床恶化 临床医生或传统的早期预警系统，从而允许及时干预。建立在我们的 在住院患者的血液恶性肿瘤亚群中的经验，这一新的努力：1)提供了一种 将较新的机器学习(ML)方法与临床信息学相结合以改进的基础 模型针对单个患者或特定子组的能力；2)评估以下各项的影响和价值 来自电子病历(EMR)的不同变量作为预测模型的一部分；以及3)扩展 对更多真实患者群体进行此方法的评估，从而深入了解翻译 这些模型已投入临床使用。因此，该项目的具体目标是： 具体目标 目的1从电子病历中识别和提取模型变量(特征)，评价不同的特征选择 不同预测准则的优化方法及其对最大似然算法的影响。 目标2开发一种ML方法，用于处理纵向的多个异步数据流 来自电子病历的信息，提供对临床恶化的实时预测。 目的3探讨临床医生和快速反应小组对早期预测临床恶化的反应。 随着这一提议的成功完成，预测模型将整合到电子病历系统中。 作为R01提案的一部分，未来的方向将涉及其他机构的外部验证和评估 对病人护理的临床影响。

Project Summary/Abstract Approximately 5-10% of hospitalized patients suffer significant clinical deterioration after admission, resulting in either transfer to the intensive care unit (ICU) or a "code" event (i.e., cardiac or pulmonary arrest). Delayed identification of these events result in increased morbidity and mortality. Unfortunately, existing prediction models result in multiple false alarms for every true positive alarm that they generate. In addition with every passing year, new monitoring systems are introduced that generate more false alarms, resulting in alarm fatigue which has been associated with patient deaths. The objective of this mentored career development proposal is to develop and assess novel computational algorithms that can predict the clinical deterioration of hospitalized patients earlier and more accurately than clinicians or conventional early warning systems, thereby allowing for timely intervention. Building upon our experience in the hematologic malignancy subpopulation of hospitalized patients, this new effort: 1) provides a foundation upon which to combine newer machine learning (ML) methods and clinical informatics to improve the capabilities of the model for an individual patient or specific subgroup; 2) assesses the impact and value of different variables from the electronic medical record (EMR) as part of the predictive model; and 3) broadens the evaluation of this approach to additional real-world patient populations, enabling insight into the translation of the models to clinical usage. The specific aims of this project are thus: Specific Aims Aim 1 To identify and extract model variables (features) from the EMR, evaluating different feature selection methods to optimize different predictive criterion and their impact on ML algorithms. Aim 2 To develop an ML approach that handles multiple asynchronous data streams of longitudinal information from the EMR, providing predictions on clinical deterioration in real-time. Aim 3 To explore clinician and rapid response team responses to early prediction of clinical deterioration. With successful completion of this proposal, the prediction model will be integrated into the EMR system. Future direction as part of a R01 proposal will involve external validation at other institutions and assessment of clinical impact on patient care.