Developing a decision model to early predict ICU admission for COVID-19 patients: A machine learning approach

Abstract

Emergency department (ED) overcrowding is a significant problem in the US. This paper develops a decision model to mitigate ED overcrowding by helping hospitals proactively plan patient boarding processes. The information obtained after the initial assessment of COVID-19 patients in the ED, including patient demographics and medical history, is utilized to predict ICU admission earlier. The predicted information can be communicated with the inpatient unit to prepare an ICU bed for the patients who need ICU care. As a result, the boarding time when patients wait for an ICU bed to be ready can be reduced. The data used in this study included 100 features and 19,155 COVID-19 patients from an academic medical center located in the Southeast United States. Multiple feature selection methods along with Extreme Gradient Boosting (XGBoost) were utilized to develop the models. The parameters of the XGBoost models are optimized using simulated annealing (SA). Among the proposed models, the best model included ten features and resulted in an area under the curve (AUC) of 89.2%, which is the highest among the models proposed in the literature. The proposed prediction model allows hospital administrators to allocate ICU beds more efficiently, enhance patient flow, and mitigate ED overcrowding.