Feature Learning for Accurate Time Prediction in Congested Healthcare Systems

Time prediction in healthcare systems such as outpatient clinics, hospital wards, and emergency departments is an essential component of decision making. Predictions are used for effective and efficient resource allocation, optimised ambulance routing, and accurate delay announcement. In this work, we focus on time prediction in congested healthcare systems, where patients share scarce resources such as nurses, physicians, and MRI machines. To achieve accurate time prediction in this setting, features describing the clinical state of the patient (e.g., severity of condition, age, and medical history) need to be combined with features that capture cross-patient information. To include the interplay of patients in time prediction, we present a method to learn congestion-related features, such as the current number of patients in the hospital and recent lengths-of-stay. To this end, we propose the model of congestion graphs, which are grounded in queueing theory and are mined from event logs as they are readily available in today’s healthcare information systems. We evaluated the proposed method on two real-world healthcare systems, namely an Israeli emergency department and an outpatient cancer hospital in the United States. Our experimental results show that using congestion feature learning we get up to 40% improvement when predicting the remaining time in the system, and 20% improvement when predicting the time to meet with the first provider.