A Framework for Evaluation and Identication of Time Series Models for Multi-Step Ahead Prediction of Physiological Signals

Time Series Modeling;Multi-Step Ahead Prediction;Physiological Signals;Mechanical Engineering;Engineering;Mechanical Engineering

El Moaqet, Hisham Mohammad Wahbi SubhiRamachandran, Satya Krishna ;

University of Michigan

关键词: Time Series Modeling;    Multi-Step Ahead Prediction;    Physiological Signals;    Mechanical Engineering;    Engineering;    Mechanical Engineering;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/116666/elmoaqet_1.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Significant interest exists in the potential to use continuous physiological monitoring to prevent respiratory complications and death, especially in the postoperative period. Smart alarm-threshold based systems are currently used with hospitalized patients. Despite clinical observations and research studies to support benefit from these systems, several concerns remain. For example, a small difference in a threshold may significantly increase the alarm rate. A significant increase in alarm related adverse outcomes has been reported by health care oversight organizations. Also, it has been recently shown that the signaled alarms are indeed late detections for clinical instability leading to a delayed recognition and less successful clinical intervention.This dissertation advances the state of art by moving from just monitoring towards prediction of physiological variables. Moving in this direction introduces research challenges in many aspects. Although existing literature describes many metrics for characterizing the prediction performance of time series models, these metrics may not be relevant for physiological signals. In these signals, clinicians are often concerned about specific regions of clinical interest. This dissertation develops and implements different types of metrics that can characterize the performance in predicting clinically relevant regions in physiological signals.In the era of massive data, biomedical devices are able to collect a large number of synchronized physiological signals recording a significant time history of a patient;;s physiological state. Directionality between physiological signals and which ones can be used to improve the ability to predict the other ones is an important research question. This dissertation uses a dynamic systems perspective to address this question. Metrics are also defined to characterize the improvement achieved by incorporating additional data into the prediction model of a physiological signal of interest.Although a rich literature exists on time series prediction models, these models traditionally consider the (absolute or square) error between the predicted and actual time series as an objective for optimization. This dissertation proposes two modeling frameworks for predicting clinical regions of interest in physiological signals. The physiological definition of the clinically relevant regions is incorporated in the model development and used to optimize models with respect to predictions of these regions.