A Study of Emergency Department Patient Admittance Predictors

We introduce and compare two prediction systems on the task of replicating human decisions regarding patient admittance in a typical American emergency department. The data-set used describes the patient trajectories in a 65,000 patient per-year emergency department in the United States. Among the descriptive attributes those of prime importance are the severity of the patient’s condition and the time they waited to be admitted from the waiting room to the department proper. A recurrent neural network (RNN) is developed to learn the task of selecting the next patient from the waiting-room/queue to be admitted for treatment which is then compared to a heuristic-based selection algorithm currently used in industry for hospital simulation applications. We demonstrate achievable accuracies of 75.29% and 84.97% using the RNN, depending on the type of the data preprocessing used. These accuracies are only potentially and theoretically achievable, respectively. The former’s validity hinges on whether certain "anomalous cases" are outliers or not, the second is achieved with the assumed existence of a method for labeling these same cases as anomalous as part of the RNN’s input, which may or may not be achievable, pending further consultation with industry experts. Our conclusions hinge on whether or not such cases are outliers though in either case a more sophisticated data-set is desired. If they are not outliers then a more detailed data-set is likely necessary to apply machine learning, or at least our methods, meaningfully to this prediction problem for use in simulated, or real world, hospitals.