# Deep Learning for Emergency Department Sustainability: Interpretable Prediction of Revisit

## TL;DR

A deep learning model predicts emergency department revisit risk using electronic health records, helping clinicians identify high-risk patients and reduce overcrowding.

## Contribution

A multimodal CNN with interpretable features using structured and unstructured EHR data to predict ED revisit risk with practical usability.

## Key findings

- The model achieved 0.717 sensitivity and 0.846 accuracy in predicting ED revisit risk.
- Binary transformation of key physiologic measurements improved model performance compared to numeric representations.
- SHAP analysis revealed unstructured physician notes dominated predictions, while structured data provided complementary insights.

## Abstract

Background: Emergency department (ED) overcrowding strains clinicians and potentially compromises urgent care quality. Unscheduled return visits (URVs), also known as readmissions, contribute to this cycle, motivating tools that identify high-risk patients at discharge. Methods: This study performed a retrospective study using ED electronic health records (EHRs) from Kaohsiung Veterans General Hospital from January 2018 to December 2022 (n = 184,653). The model integrates structured variables, such as vital signs, medication and laboratory counts, and ICD-10–based comorbidity measures, with unstructured physician notes. Key physiologic measurements were transformed into binary form using clinical reference intervals, and random under-sampling addressed class imbalance. A multimodal, CNN was proposed and evaluated with an 8:2 train–test split and 10-fold Monte Carlo cross-validation. Results: The proposed model achieved a sensitivity of 0.717 (CI: [0.695, 0.738]), accuracy of 0.846 (CI: [0.842, 0.850]), and AUROC of 0.853. Binary transformation improved recall and AUROC relative to the original numeric representations. SHAP analysis showed that unstructured features dominated prediction, while structured variables added complementary value. In a small-scale pilot evaluation using the SHAP-enabled interface, participating physicians reported the system helped surface high-risk cohorts and reduced cognitive workload by consolidating relevant patient information for rapid cross-checking. Conclusions: An interpretable CNN-based clinical decision support system can predict ED revisit risk from multimodal EHR data and demonstrates practical usability in a real-world clinical setting, supporting targeted discharge planning and follow-up as a near-term approach to mitigate overcrowding.