Optimal Patient-centered Response to Acute Physiological Deterioration of Hospitalized Patients.

摘要

Hospitalized patients are at risk of unexpected acute and persistent physiological deterioration (APD), which is identified by disturbance in one or multiple physiological measures. Unanticipated APD may result in respiratory instability, cardiopulmonary arrest and death. Early warning scores (EWSs) are recommended as part of the early recognition and response to inpatient deterioration, including APD. EWSs not only provide a standardized method for clinical assessment, they also suggest which patients may require attention outside of an intensive care unit (ICU) provided by a critical care team, also called a Rapid Response Team (RRT). Currently used EWSs differ in the included physiological measures, and in the thresholds for triggering a response, e.g., RRT activation. At this time, there is no consensus on clinical guidelines for selection of physiological measures, and their thresholds to inform acute care decisions. There is a growing need to use the EWSs for acute care decision making to avoid failed or delayed response to APD. Bedside providers in the general ward commonly have to rely on subjective evaluations to trigger a response to APD. In addition, the current use of EWSs relies on fixed thresholds for a response without considering patient characteristics. Exploring the relationship between patient characteristics and APD, and identifying the patients who may benefit from an increased level of acute medical care would provide guidance in RRT activation, and help to personalize acute medical care.;In this research, we collaborate with the Division of Health Care Policy & Research, Department of Health Sciences Research, Mayo Clinic, Rochester. We seek to individualize the response to APD by using electronic medical records (EMRs). We use EWSs to capture the stochastic changes in a patient's physiological condition during a stay in the general ward. Our methodology includes statistical analysis, dynamic programming, random variable generation, clustering analysis, and a robust modeling approach. We segment the data into patient subpopulations, and apply the Chi-square and Kruskal-Wallis tests to identify statistically significantly different subpopulations. We develop subpopulation-specific infinite-horizon semi-Markov decision process (SMDP) models to optimize the care metrics related to stabilization and failure to recognize APD, while capturing the provider resource use as a function of EWSs and RRT activation. The optimal policies identify the subpopulation-specific RRT thresholds. We provide theoretical insights into the optimal total expected costs, and identify the framework to prove the existence of a control-limit policy. Finally, we address the uncertainty in cost parameters, because the fixed costs include time-based provider resource use, and may be subject to errors. In addition, the same resource time may be valued differently depending on the providers' expertise level and the patients' needs. We use an experimental design with eight scenarios, combined with a robust SMDP framework, to explore the impact of uncertainty in costs on the model results. The results of this research will allow bedside providers to make informed decisions regarding triggering an individualized response to APD in the general ward, and provide a baseline for future research in acute care decision making.