Deep Learning for Emergency Department Sustainability: Interpretable Prediction of Revisit

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Medicine
• Health Services Research

Background:

• Emergency department (ED) overcrowding is a significant issue impacting patient care quality.
• Unscheduled return visits (URVs) exacerbate ED overcrowding, necessitating predictive tools for patient discharge.
• Identifying patients at high risk for URVs is crucial for effective resource allocation and patient management.

Purpose of the Study:

• To develop and evaluate a multimodal deep learning model for predicting the risk of unscheduled return visits to the emergency department.
• To integrate structured and unstructured electronic health record (EHR) data for enhanced predictive accuracy.
• To assess the clinical utility and interpretability of the predictive model in a real-world setting.

Main Methods:

• A retrospective analysis of 184,653 EHRs from Kaohsiung Veterans General Hospital (2018-2022).
• Development of a multimodal Convolutional Neural Network (CNN) integrating structured (vitals, labs, meds, comorbidities) and unstructured (physician notes) data.
• Physiologic measurements were binarized, and class imbalance was addressed using random under-sampling; model evaluation employed an 8:2 train-test split and 10-fold Monte Carlo cross-validation.

Main Results:

• The CNN model achieved a sensitivity of 0.717, accuracy of 0.846, and AUROC of 0.853.
• Binary transformation of physiologic data improved model recall and AUROC.
• SHAP analysis indicated that unstructured physician notes were primary predictors, with structured data offering complementary value; pilot evaluation showed reduced cognitive workload for physicians.

Conclusions:

• An interpretable CNN-based clinical decision support system effectively predicts ED revisit risk using multimodal EHR data.
• The system demonstrated practical usability, aiding clinicians in identifying high-risk patients for targeted discharge planning.
• This approach offers a viable strategy to mitigate ED overcrowding by enabling proactive patient management.