Deep learning to predict emergency department revisit using static and dynamic features (Deep Revisit): development...

Area of Science:

• Emergency Medicine
• Artificial Intelligence
• Clinical Informatics

Background:

• Emergency Department (ED) revisits are a significant concern, with high-risk revisits requiring urgent attention.
• Existing machine learning models for ED revisit prediction often underutilize dynamic patient features.
• Deep learning approaches for this problem are relatively unexplored.

Purpose of the Study:

• To develop and evaluate a novel hybrid deep learning model for predicting Emergency Department revisits.
• To integrate both static and dynamic clinical features for improved prediction accuracy.
• To identify high-risk ED revisit cases more effectively.

Main Methods:

• A hybrid deep learning model combining Temporal Convolutional Network (TCN) and FT-Transformer was developed.
• The model utilized static (age, sex, triage) and dynamic (vital signs) features from National Taiwan University Hospital (NTUH) data.
• A preprocessing strategy was implemented to handle temporal data irregularities.

Main Results:

• The model achieved an AUROC of 0.8453 for high-risk revisits and 0.7250 for general revisits.
• Compared to a static-only logistic regression baseline, the hybrid model showed substantial improvements in AUPRC and precision.
• The model demonstrated robust performance on validation data from different time periods.

Conclusions:

• The proposed hybrid deep learning model significantly outperforms traditional methods for ED revisit prediction.
• Multimodal clinical data fusion using deep learning is effective for enhancing ED revisit prediction.
• The model shows promise in supporting clinical decision-making for patient management.