Unobtrusive neonatal patient monitoring using video and pressure data

我们的研究旨在通过开发新的不显眼的传感器和算法来改善新生儿重症监护病房(NICU)的患者护理。在过去的2.5年里，在NSERC CRD的支持下，我们从东安大略省儿童医院(CHEO)的NICU收集了独特的、丰富的多传感器患者数据集。通过与IBM高级研究中心和CHEO临床医生的持续合作，我们现在将使用机器学习和计算机视觉来挖掘这些数据，以创建新型非接触式患者监控系统，以增强NICU的患者护理。我们最初的NSERC CRD赠款资助收集了35名新生儿的独特数据集，这些数据跨越了不同的患者群体和护理水平(床位类型)。数据收集自压敏垫(PSM)--首次用于新生儿；英特尔RealSense RGB-D相机，用于测量颜色、近红外和深度数据；黄金标准生理数据，来自带有有线电极的患者监护仪。此外，我们还开发了一款定制的平板电脑应用程序，以便床边的研究人员能够在收集传感器数据的同时注释所有临床感兴趣的事件。这些数据是独一无二的，因为我们有多个传感器数据流与黄金标准生理数据和临床事件注释同时收集。本研究将开发和评估机器学习和计算机视觉算法，以估计来自非接触式传感器的生理信号，如呼吸频率。我们的时态事件注释(TEA)框架将作为开源项目进行扩展和发布，以使其他研究小组能够仅基于要捕获的事件类型的描述来自动生成原生数据收集应用程序。最后，还将检测和表征患者的运动，最终目标是减少运动伪影造成的错误警报。

Our research aims to improve patient care in the neonatal intensive care unit (NICU) through the development of novel unobtrusive sensors and algorithms. Over the past 2.5 years, with NSERC CRD support, we have collected a unique, rich, multi-sensor patient dataset from the NICU at the Children's Hospital of Eastern Ontario (CHEO). Through continued collaboration with IBM's Centre for Advanced Studies and clinicians at CHEO, we will now mine these data using machine learning and computer vision to create novel non-contact patient monitoring systems to augment patient care in the NICU. Our original NSERC CRD grant funded the collection of a unique dataset from 35 neonates spanning different patient masses and levels of care (bed types). Data were collected from a pressure-sensitive mat (PSM) - a first for the neonatal population; an Intel RealSense RGB-D camera that measures colour, near-infrared, and depth data; and gold-standard physiologic data from the patient monitor with wired electrodes. In addition, we have developed a custom tablet app such that a bed-side researcher was able to annotate all events of clinical interest simultaneously with the sensor data collection. These data are unique, in that we have multiple sensor data streams collected simultaneously with gold standard physiologic data and clinical event annotations. The present research will develop and evaluate machine learning and computer vision algorithms for estimating physiologic signals, such as respiration rate, from non-contact sensors. Our Temporal Event Annotation (TEA) framework will be extended and released as an open-source project to enable other research groups to automatically generate native data collection apps based solely on a description of the event types to be captured. Lastly, patient movement will also be detected and characterized, with the ultimate goal to reduce false alarms resulting from motion artifacts.