Prediction of Hospital Readmission for Heart Disease: A Deep Learning Approach

Abstract

Hospital readmissions consume large amounts of medical resources and negatively impact the healthcare system. Predicting the readmission rate early one can alleviate the financial and medical consequences. Most related studies only select the patient’s structural features or text features for modeling analysis, which offer an incomplete picture of the patient. Based on structured data (including demographic data, clinical data, administrative data) and medical record text, this paper uses deep learning methods to construct an optimal model for hospital readmission prediction, tested on a dataset of heart disease patients’ 30-day readmission. The results show that when only structured data is used, the deep learning model is much better than the Naive Bayes model and slightly better than the Support Vector Machine model. Adding a text model to the deep learning model improves performance, increasing accuracy and F1-score by 2% and 6%, respectively. This indicates that textual information contributes greatly to hospital readmission predictions.