2021 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI) | 2021
Applying Problem Transformation Methods for Predicting 3-Days All-Cause Readmission in the Emergency Department
Abstract
For an emergency department (ED), patient readmissions after ED discharges in a short period are usually considered as an indicator to evaluate the quality of medical services, representing whether the patients had received appropriate treatments. Moreover, these unplanned revisits may cause ED crowding, leading frustrating patient experiences. In order to prevent readmission, many studies had tried to build predictive models for the further intervention for high-risk populations. However, the binary prediction of readmission is much harder than other tasks. Usually, predicting readmission is binary classification task, and people intuitively build a single model to solve the problem. Yet, as the purpose of multi-label learning, that may be helpful for model performance by providing additional helpful information between these related labels. Specifically, this study extends the prediction of readmission to a multi-label form through assembling another two related labels, then applies problem transform methods for pursuing a better prediction on the original task. For verifying the idea, four empirical tree-based machine learning methods were trained and evaluated with more than four 40,000 ED admission records, either in a simple binary classification form or a multi-label form. The results demonstrate a notable improvement of F1-scores and precision for predicting readmissions through a problem transformation pipeline. The improvement is attributed to the proposed pipeline that utilizes the mutual information for classifying through gathering additional related labels. In this study, we provided alternative solution for the further studies that aim to improve the performance of binary classification with a potential mutual related dataset.