Imbalanced learning to predict long stay Emergency Department patients

A major contributor to Emergency Department (ED) crowding is patients with prolonged length of stay (LOS). Patients with long stays (i.e., those with LOS longer than 14 hours) comprise 10% percent of ED visits, but utilize 30% of the total ED bed hours. Accurately predicting patients' LOS can be used to improve resource management both in the ED and the hospital. A prediction model that can identify this minority, prolonged stay patient group, early at presentation may be effective in addressing barriers to expedited treatment and ED disposition. However, this is a challenging task because regular classification techniques are biased toward the majority group of examples and tend to overlook the minority class examples. This problem can be alleviated by using class imbalance learning methods. In this paper, we present a framework that predicts patients with prolonged ED stays (> 14 hours) from data available at triage (i.e., presentation). The framework also enables extraction of independent variables that capture the current state of the resources in the ED. Predictions combine patient information (e.g., demographics, complaints, and vital signs) with a snapshot of resources and queuing metrics in the ED which can substantially impact the LOS. The prediction models in our framework are developed from over one hundred thousand ED encounters retrospectively collected at an urban hospital. Our experimental results demonstrate that we accurately predict prolonged ED length of stay and provide a clear interpretation of the factors that influence it. We also found that integrating a class imbalance learning ensemble method into our framework produces much better results for prolonged stays than only using traditional logistic regression methods.