Develop&validate SuperAlarm to detect patient deterioration with few false alarms

 描述(由申请人提供)：重症监护病房(ICU)的多参数患者监测仍然不令人满意，众所周知的警报疲劳问题就是明证。据报道，一名重症监护病人每天可以发出700个警报。在加州大学旧金山分校，在一个月的评估期间，ICU每天平均每张病床发生187次声音警报。改进信号处理算法，微调警报阈值，将一些警报降级到听不见的类别，是解决警报疲劳的一些方法。然而，这些干预措施试图在传统的患者监测实践的背景下解决问题(即，重点放在单个警报上，而忽略警报之间的关系、由电子健康记录(EHR)中的其他数据建立的上下文信息以及所有这些变量的顺序模式)。事实上，我们认为，警报疲劳反映了重症监护临床医生面临的更深层次的挑战，他们被越来越可用的原始数据超载，但没有适当的工具来利用这些数据的潜力来治疗他们的患者。作为支持临床医生克服数据过载的第一步，我们的目标是准确地 通过识别单个警报、生理变量和EHR数据之间的组合和顺序模式，检测患者的总体状态变化。实现这一目标将导致开发额外的决策支持工具，以了解原因并为检测到的患者状态变化选择潜在的干预措施。我们小组已经进行了初步研究，证明了实现这一目标的可行性。特别是，我们已经发展了一种特定的算法来识别共发生的监测警报，这种警报经常发生在医院内的心肺骤停(CPA)之前，但很少发生在对照组患者中，到了数据融合框架。这个框架能够识别更丰富的变量集的预测性组合，包括实验室测试和监视器无法获得的其他生理变量。我们称这些组合为超级警报模式。通过构建，SuperAlarm触发器发生的频率将比监视器警报低得多，但在检测患者状态变化方面更准确。因此，本应用程序的目标是使用预期数据在SuperAlarm数据融合框架下开发和验证进一步的算法改进。我们将追求以下三个目标：1)通过新的心电信号分析来丰富SuperAlarm模式；2)发展SuperAlarm触发序列的序列模式识别方法。3)进行前瞻性数据收集，以开发和验证SuperAlarm模型。

 DESCRIPTION (provided by applicant): Multi-parameter patient monitoring in intensive care units (ICU) remains unsatisfactory as evidenced by the well-known alarm fatigue problem. It has been reported that a critical care patient could generate 700 alarms per day. At UCSF, a daily average of 187 audible alarms per bed occurred in ICUs during one month of assessment. Improving signal processing algorithms, fine-tuning alarm thresholds, and downgrading some alarms to an inaudible category are some ways to address alarm fatigue. However, these interventions attempt to solve the problem within the context of conventional patient monitoring practice (i.e., the focus is on an individual alarm while ignoring the relationships among alarms, the contextual information established by other data available in electronic health record (EHR), and the sequential patterns of all of these variables). In fact, we argue that alarm fatigue reflecs a deeper challenge for critical care clinicians who are overloaded with increasingly available raw data but do not have appropriate tools to leverage the potential of these data to treat their patients. As a first step to support clinicians to overcome data overload, our goal is to precisely detect gross patient state changes by recognizing combinatorial and sequential patterns among individual alarms, physiological variables, and EHR data. Achieving this goal will lead to developing additional decision support tools to understand the causes and select potential interventions for the detected patient state changes. Our group has done preliminary studies that demonstrate the feasibility of achieving this goal. In particular, we have evolved a specific algorithm to identify co-occurring monitor alarms, which frequently precede in-hospital cardiopulmonary arrests (CPA) but rarely occur among control patients, to a data fusion framework. This framework is capable of recognizing predictive combinations of a much richer set of variables including lab tests and additional physiological variables not available from monitors. We term these combinations SuperAlarm patterns. By construction, SuperAlarm triggers will occur much less frequently than monitor alarms, yet be more precise in detecting patient state changes. Thus, the objective of this application is to develop and validate further algorithm improvement under this SuperAlarm data fusion framework using prospective data. We will pursue the following three aims: 1) To enrich SuperAlarm patterns by novel analysis of Electrocardiographic (ECG) signals; 2) To develop sequential pattern recognition methods for sequences of SuperAlarm triggers. 3) To conduct prospective data collection to develop and validate SuperAlarm model.