JMIR Publications

ABSTRACT

Background:

In-hospital cardiac arrest (IHCA) is a severe and sudden medical emergency, characterized by the abrupt cessation of circulatory function, leading to death or irreversible organ damage if not addressed immediately. IHCA affects 0.1% to 0.5% of emergency department (ED) patients, and the mortality rate remains high even with advanced medical interventions. Early detection is crucial, yet identifying subtle signs of cardiac deterioration is challenging. Traditional IHCA prediction methods rely heavily on structured vital signs; unstructured 12-lead electrocardiogram (ECG) at ED triage may provide more information for early prediction of imminent ED cardiac arrest. Image-based methods can be integrated seamlessly into clinical workflows to aid healthcare professionals in making quicker and more suitable for frontline emergency settings.

Objective:

This study aims to address the challenge of early detection of in-hospital cardiac arrest (IHCA) in the ED by developing an innovative deep learning model, the ECG-Image-Aware Network (EIANet), that leverages 12-lead ECG images for prediction. Traditional methods rely on raw. single-lead ECG signals, which may require monitoring devices and signal processing before analysis. By focusing on 12-lead ECG images that can be obtained at ED triage, this research seeks to create a more accessible solution that can seamlessly integrate into real-world ED triage environments.

Methods:

For adult patients with ED cardiac arrest (cases), their 12-lead ECG images at ED triage were obtained from two independent datasets: National Taiwan University Hospital (NTUH) and Far Eastern Memorial Hospital (FEMH). Control ECGs were randomly selected from adult ED patients without cardiac arrest during the same study period. In EIANet, ECG images were first converted to binary form, followed by noise reduction using Gaussian blurring, connected component analysis, and morphological opening. A spatial attention module was incorporated into the ResNet-50 architecture to enhance feature extraction, and a custom Binary Recall Loss (BRLoss) was employed to balance precision and recall, addressing slight dataset imbalance. The model was developed and internally validated on the NTUH-ECG dataset and was externally validated on an independent FEMH-ECG dataset. The model performance was evaluated by F1-score, area under the receiver operating curve (AUROC), and area under the precision recall curve (AUPRC).

Results:

There were 571 case and 826 control ECGs in the NTUH dataset; 378 case and 713 control ECGs in the FEMH dataset. The novel EIANet model achieved an F1-score of 0.805, an AUROC of 0.896, and an AUPRC of 0.842 on the NTUH-ECG dataset with a 40% positive sample ratio. It achieved an F1-score of 0.650, an AUROC of 0.803, and an AUPRC of 0.678 on the FEMH-ECG dataset with a 34.6% positive sample ratio. The feature map showed that the region of interest in the ECG was ST segment.

Conclusions:

The EIANet demonstrates promising potential for accurately predicting IHCA in the ED using triage ECG images, offering an effective solution for early detection of high-risk cases in emergency settings. This approach may enhance the ability of healthcare professionals to make timely decisions, with potential of improving patient outcomes by enabling earlier interventions for IHCA.