Patient Length-of-Stay Prediction and Inpatient Bed Forecasting in UNC Hospitals’ Emergency Department

Planning for future inpatient arrivals originating from the emergency department (ED) can improve bed management in the hospital and coordination between the ED and inpatient units. While predicting whether or not a given patient will be admitted to the hospital is an important task, there is also utility in knowing when exactly the need for an in-patient bed will arise for patients who are ultimately admitted to the hospital. Given this knowledge, healthcare providers can take measures to speed up the discharge process or otherwise shorten the time that a patient has to wait for an inpatient bed in the ED. In light of this insight, in the first chapter of this thesis, we apply a random forest model to predict patient length-of-stay in the emergency department, defined as time from the end of triage to the time of the disposition decision. In the second chapter, we combine the outputs of the aforementioned model and another model that predicts the probability of hospital admission for any given patient. By aggregating these predictions across patients, we create retroactive forecasts of expected inpatient bed demand using data from UNC Hospitals’ emergency department. Finally, we leverage finite-sample methods from the area of conformal prediction to construct prediction intervals for these expected bed demand forecasts. These intervals provide approximately valid marginal coverage experimentally and offer a way to quantify and communicate uncertainty in bed forecasts for clinicians.