Tuesday, October 25, 2011: 1:45 PM
Columbus Hall C-F (Hyatt Regency Chicago)
(ESP) Applied Health Economics, Services, and Policy Research

Candidate for the Lee B. Lusted Student Prize Competition

Valerie Chase, Mariel S. Lavieri, PhD, Amy Cohn, PhD and Tim Peterson, MD, University of Michigan, Ann Arbor, MI

Purpose:  We investigate the use of statistical models to identify surges in emergency department (ED) volume based on the level of utilization of physician capacity. Our models may be used to guide staffing decisions in non-crisis related patient volume increases.

Method: Patient visits to a large urban teaching hospital with a Level 1 trauma center were collected from July 2009 – June 2010. A comparison of significance was used to assess the impact of multiple variables on the state of the ED. Historical physician utilization data was used to model physician capacity. Binary logistic regression analysis was used to predict the probability that the physician capacity would be sufficient to treat all patients forecasted to arrive.  The predictions were performed by various time intervals: 15 minutes, 30 minutes, 1 hour, 2 hours, 4 hours, 8 hours and 12 hours.  The models were validated against 5 consecutive months of similar patient data from July – November 2010.  Models and forecast accuracy were evaluated by positive predictive values, Type I and Type II errors, and real-time accuracy in predicting non-crisis surge events. 

Result: The ratio of new patients to treat to total physician capacity - termed the “Care Utilization Ratio (CUR)” - was deemed to be a robust predictor of the state of the ED (with a CUR ratio greater than 1 indicating that the physician capacity is not sufficient to treat all patients forecasted to arrive). Among the models investigated, prediction intervals of 30 minutes, 8 hours and 12 hours performed best with deviances of 1.000, 0.951 and 0.864 respectively. The models were validated against the July – November 2010 data set using significance of 0.05. For the 30-minute prediction intervals, the positive predictive values ranged from 0.738 to 0.872,  true positives ranged from 74% to 94%, and true negatives ranged from 70% to 90% depending on the threshold used to determine the state of the ED.

Conclusion:  We identified a new and robust indicator of the system’s performance: CUR. By investigating different prediction intervals, we were able to model the tradeoff of longer time to response versus shorter but more accurate predictions. Our proposed models would’ve allowed for an earlier identification of surge in patient volume on “non-crisis” days than current practice.