Predicting Frequent Emergency Department Visitors using Adaptive Ensemble Learning Model

BackgroundPredicting accurately the frequent Emergency Department (ED) visitors is critical for hospitals because they often consume significant ED resources, including staff time, equipment, and medical supplies. Furthermore, frequent ED visitors may contribute to increased wait times for all patients. Therefore, by accurately predicting and identifying these individuals, hospitals can help reduce the burden on the ED and decrease wait times for all patients, improving the overall quality of care. ObjectiveThis study proposed an effective and adaptive ensemble learning prediction model to identify frequent visitors in the emergency department. MethodsThis was a retrospective population-based study of patients and utilised medical and administrative databases at Canberra Hospital, a tertiary public hospital in ACT, Australia, between January 1997 and December 2022. The study focuses on a wide age range of the population with 20 viral chronic diseases. The definition of frequent ED use is considered as having at least three visits within a year. This study developed an Adaptive ensemble learning-based prediction model and compared the performance with 16 popular machine learning models. In addition, three techniques are compared to handle the imbalanced data issue, and we also proposed a hybrid feature selection composed of Elastic-Net and local search to find the best combination of features. In order to hyper-parameter tuning, two techniques were compared: a population-based evolutionary algorithm and a local search. ResultsThe study included 535,474 patient visits and 1.6 million episodes, with 25% overall frequent visitors. We compared the performance of the proposed prediction model with that of the other 16 popular classifiers. According to the prediction results, the proposed model considerably outperformed other models in terms of five metrics: accuracy, Recall, F1-score, Area under the ROC curve (AUC), and Log loss at 0.78 (95% CI 0.78-0.79), 0.68 (95% CI 0.68-0.68), 0.68 (95% CI 0.68-0.69), 0.69 (95% CI 0.69-0.70), and 7.4 (95% CI 7.2-7.5), respectively. ConclusionsWe proposed an adaptive ensemble learning model combining XGBoost Elastic-net with local search and Differential evolution to address the imbalanced nature of the frequent ED visitors data. Our approach aimed to enhance the prediction capability of the classifier substantially. To tackle the class imbalance, we employed both under-sampling and adjusted weights for the positive class. Through extensive testing and evaluation, we demonstrated that these strategies effectively improved the models performance. Further, we emphasised the importance of employing a robust feature selection method and a fast hyper-parameter optimiser. These elements were essential for enhancing the identification of frequent ED visitors. By incorporating these techniques, our study contributes to developing more accurate and reliable models for predicting frequent ED visitors, thereby assisting hospitals in resource allocation and patient care management.