Real-time predictive modeling methods using multi-rate physiological time series data for the prediction of disease onset

这项研究计划的主要目标是加速最先进的预测建模方法的发展，以便预测疾病的发生和其他情况。这些方法将结合多个多速率生理时间序列数据流，包括：快速数据(通常以每秒250-1000个样本测量)，例如心电图、阻抗呼吸波形和静脉血压；低速数据，例如血氧饱和度；以及其他临床信息。目前的建模方法仅限于对这些波形的简单分析，这排除了除了临床医生自己之外考虑它们之间的相互作用。 程序中使用的实时预测建模方法将系统辨识方法(如快速正交搜索)与预测方法(如用于非线性模型预测控制的方法)相结合。快速正交搜索提供了一种有效地组合多速率数据源和确定简约模型的方法。 最初，这些方法将在两种情况下应用。第一个是新生儿重症监护病房，目前正在通过一项名为Artemis项目的合作试点研究收集数据流，该研究涉及UOIT、多伦多患病儿童医院和IBM的TJ Watson研究实验室。第二种是通过与护理机构的另一项合作实现居家护理。 由拟议研究产生的模型和新方法将可用于临床研究和在现实世界中实施。这些新的实时预测建模方法在实时事件流处理至关重要的领域有潜在的无限应用，特别是在将启用自主决策和控制的领域。

The primary objective of this research program is to accelerate the development of state of the art predictive modeling methods in order to anticipate the onset of disease and other conditions. These methods will incorporate multiple multi-rate physiological time series data streams, including: fast rate data (typically measured at 250-1000 samples per second), such as electrocardiogram, the impedance respiratory waveform and intravenous blood pressure; slow rate data, such as blood oxygen saturation; and other clinical information. Current modeling methods are limited to simple analysis of these waveforms which precludes the consideration of their interactions except by clinicians themselves. The real-time predictive modeling methods to be used in the program combine system identification methods (such as Fast Orthogonal Search) with predictive methods such as those used in nonlinear model predictive control. Fast Orthogonal Search provides a means of effectively combining multi-rate data sources and determining parsimonious models. Initially, these methods will be applied in two contexts. The first is the neonatal intensive care unit where data streams are currently being collected in a collaborative pilot study involving UOIT, the Hospital for Sick Children in Toronto, and IBM's TJ Watson Research Laboratory, called the Artemis project. The second is home care through another collaboration with a care facility. The models and new methods resulting from the proposed research will be made available for use in clinical studies and implementation in the real world. There are potentially unlimited applications for these new real-time predictive modeling methods in areas where real-time event stream processing are critical, particularly where autonomous decision making and control will be enabled.