,

Principal Investigator: Thomas Kirsch, MD, MPH and Scott Levin, PhD

Research Question:

The objective is to improve hospital medical surge capacity by developing a rapid and accurate decision support tool to predict appropriate level of in-hospital care. This objective stems from the hypothesis that patient information available in the emergency department (ED) may be used to predict appropriate disposition of ED patients. ED patients triaged to inadequate levels of care are placed at higher risk. ED patients triaged to excessively high levels of care receive no added benefit and unnecessarily consume limited critical care resources. A decision support tool has potential to quickly evaluate the associated risk of triaging patients to different levels of care. Improved triage decisions will lead to a lower rate of adverse outcomes, reduced length-of-stay, and improved allocation of scarce critical care resources. Implementation of a triage tool would be beneficial during normal hospital operations, but particularly useful during disaster settings when swift decision making, resource allocation and capacity management is vital.

This project meets major healthcare disaster preparedness and response goals established in HSPD-21, ESF-8 of the National Response Framework and the work of both the Department of Homeland Security (DHS) Office of Health Affairs (OFA) and the Department of Health and Human Services (HHS). It also meets some of the goals of other federal preparedness and response projects by using novel information management and decision-support tools. The project will address the goals of improving mass casualty care, disaster health systems and healthcare surge capacity as well as identifying evidence-based standards of care emphasized by HSPD-21, DHS OHA and by HHS. This tool will improve mass casualty care by immediately re-orienting and coordinating hospital resources with pre-hospital needs to increase surge capacity by distributing hospital beds optimally.

Analytic Approach:

The tool will utilize physiologic, historic health, demographic and ED treatment information (e.g., diagnoses, treatments, procedures) to predict outcomes associated with assigning patients to alternative levels of care. Outcome variables to be assessed include: in-hospital mortality; transitions to higher levels (e.g., patient transfers from an inpatient medicine unit to an intensive care unit); adverse events defined by the Agency for Healthcare Research in Quality (AHRQ); length of stay; and hospital discharge within 24 hours. A more detailed description of the predictor and outcome variables may be seen in Data Collection (Section 3) below.

The research design will be a multi-site retrospective cohort study of ED admissions through medicine services over a 2 year time-period. Patient information will be collected from each of the three sites which include Johns Hopkins Hospital (original site), Vanderbilt Medical Center, and Nashville General Hospital at Meharry. Patients within the cohort at each site will be stratified by admission care level (i.e., intensive care, intermediate care, floor care). Outcomes and specific ED indicators hypothesized to predict outcomes will be extracted electronically from multiple clinical information systems.

The relationship between patient predictor variables collected in the ED and outcomes stratified by level of care will be examined using a two-phase logistic regression approach. This approach will be used to determine odds-ratios of outcomes associated with assigning patients to each care-level. The first phase of model development involves determining which ED variables currently predict care-level for admitted patients. Two logistic regression model structures will be applied concurrently and compared to predict care-level. A pairwise and ordinal logistic regression approach will be used to predict care-level as the dependent (i.e., outcome) variable. The second phase of the analysis determines the appropriate care-level assignment based on outcomes. This is done by using a propensity score matching strategy to correct for selection bias. Patient propensity scores (i.e., probability of care-level assignment) and potentially other predictor variables (e.g., diagnosis) will be used to create sub-groups of matching patients via cluster analysis. These sub-groups will be analyzed using identical regression models where aforementioned outcomes are the dependent variable. Independent variables are “potential” care-level and any propensity matching predictor variables (e.g., propensity score). Care-level coefficients may be examined to determine the strength and significance of care-level assignment in predicting outcomes. Collectively evaluating care-level coefficients for each outcome guides decisions of appropriate assignment. It is important to note that modeling methods described may be applied to any hospital which has more than one inpatient care level. For, example models are easily adaptable to hospitals with no intermediate care by simply dropping the intermediate care model variable.

Logistic regression models used in the final decision support tool will be continuously validated as the proposed research shifts from a single to multi-site study. Cross-validation using a training set (2/3 data) and testing set (1/3 data) at the original site will be conducted. External validation and analysis re-creation using a more limited set of predictor variables will be conducted at the two alternate sites.