Deep Learning-Based Electrocardiogram Model (EIANet) to Predict Emergency Department Cardiac Arrest: Development and...

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Emergency Medicine

Background:

• In-hospital cardiac arrest (IHCA) is a critical emergency; emergency department cardiac arrest (EDCA) accounts for 10-20% of cases.
• Early EDCA detection is vital but challenging due to subtle deterioration signs.
• Traditional prediction methods rely on structured vital signs or ECG signals, requiring extra processing or devices.

Purpose of the Study:

• To develop a deep learning model (EIANet) for early EDCA prediction using 12-lead ECG images from emergency department triage.
• To create a practical and accessible solution for seamless integration into clinical workflows.

Main Methods:

• A novel ECG-Image-Aware Network (EIANet) was developed using ResNet50 with a spatial attention module.
• ECG images underwent preprocessing including binarization, noise reduction, and morphological opening.
• The model was trained and validated on independent datasets (NTUH, FEMH) using custom binary recall loss (BRLoss).

Main Results:

• EIANet achieved high performance: F1-score 0.805, AUROC 0.896, AUPRC 0.842 on the NTUH dataset.
• External validation on the FEMH dataset showed F1-score 0.650, AUROC 0.803, AUPRC 0.678.
• Feature analysis identified the ST segment as the region of interest for prediction.

Conclusions:

• EIANet shows significant potential for accurate EDCA prediction using readily available triage ECG images.
• This AI-driven approach can enhance early detection of high-risk patients in emergency settings.
• Timely decision-making and earlier interventions may improve patient outcomes for EDCA.