A deep learning approach for predicting early bounce-backs to the emergency departments

Reviewing patients who return to the emergency department (ED) within 72 h (i.e., bounce-back) is a standard quality assurance procedure used to identify correctable system- and clinician-level causes for earlier-than-expected return to the ED and ultimately ensure patients’ safety. This study proposes a deep learning (DL) framework to automatically extract features from structured and unstructured Electronic Health Records (EHR) data of ED visits and predict patients who are likely to bounce-back. Data from 120,000+ visits to the ED of four major hospitals in New York city over 4+ years are used to validate the proposed framework. The DL model was able to predict 74.8% of the bounce-backs (AUROC= 0.766). Notably, our results show that leveraging state-of-the-art DL techniques to extract features from unstructured ED physician narrative notes and incorporating them (along with structured vitals and demographics data) in the prediction task can remarkably improve the outcome. In addition, a sensitivity analysis performed on the structured features of the model revealed that factors such as patients’ age, type of medical insurance, chronic conditions (especially renal, heart, and respiratory), heart failure, and head or neck trauma are among the top factors increasing the chances of patients’ unplanned return to the ED. The proposed framework may be used as a decision support tool to assist emergency care clinicians in the early identification of the high-risk patients who are likely to bounce-back and provide them with timely, appropriate care. •Bounce-back to the Emergency Department is a serious healthcare quality issue.•Physicians’ narratives involve valuable information to help predicting bounce-backs.•Deep learning algorithms are used to extract crucial patterns from medical narratives.