811 - Predictive Modeling for Daily Emergency Department Visits
Friday, May 15, 2020
Location: Majestic Ballroom: Majestic
Background and Objectives: The Froedtert Emergency Department (ED) is the only level 1 trauma center in southeastern, Wisconsin with a high average daily volume of 210 patients per day. To assist with daily hospital and departmental capacity constraints, resources need to match daily demand. Forecasting ED volume can help inform ED staffing models and hospital bed availability. The objective was to create a simple predictive model that would forecast daily ED volume at the Froedtert ED.
Methods: Daily ED volumes from January 1st, 2016 through December 31st, 2017 were collected and recorded using the electronic medical record system, Epic (Verona, WI). Indicator variables were collected and evaluated including day of the week (Sunday through Saturday), specific national holidays (Thanksgiving, Christmas eve, Christmas day, New Year’s Eve, New Year’s Day, Fourth of July, Memorial Day, and Labor day), day after listed holidays, relevant daily lags, and daily average temperature. The response variable used was daily ED volume. A time series regression analysis was then performed using JMP Statistical Software (Cary, NC).
Results: Daily volume at Froedtert ED was found to be a function of Monday, Saturday, holiday, holiday + 1 day, lag - 1 day, and average daily temperature. The model produced: Daily ED Volume = 155.3465 + 21.983429(Monday) – 5.575067(Saturday) – 22.97639(Holiday) + 21.56593(Day after Holiday) + .2159412(Temp) + .1639083(Lag-1). The R-squared value for the model was 31.7.
Conclusion: Accurately predicting daily ED volumes has potential benefits for both department staffing and hospital capacity planning, however, given the underlying complex human behavior elements of an ED visit, creating an accurate model is very challenging. While this model’s R-squared value was only modest for accuracy and predictability, from a practical standpoint a model does not need to accurately predict to a singular patient. Future studies could create a similar model and then develop a classification tree for probabilities of volumes above or below critical thresholds. Additional indicator variables could also be investigated for correlation and future model accuracy.