Image based deep learning in 12-lead ECG diagnosis.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Electrocardiograms (ECGs) are crucial for diagnosing cardiovascular diseases.
• Current machine learning ECG studies often use raw signal data, limiting applicability when only ECG images are available.
• Accessibility of ECG images is a challenge in remote and regional healthcare settings.

Purpose of the Study:

• To evaluate the diagnostic accuracy of image-based deep learning algorithms for 12-lead ECGs.
• To assess the feasibility of using deep learning on ECG images for clinical applications.

Main Methods:

• Deep learning models with VGG architecture were trained on diverse 12-lead ECG datasets.
• Model performance was evaluated using holdout test data and data from unseen datasets.
• Grad-CAM was employed to visualize diagnostic feature heatmaps on ECG images.

Main Results:

• The models achieved excellent accuracy (AUROC, AUPRC, sensitivity, specificity) on test data.
• Performance was comparable to leading signal and image-based ECG diagnostic models.
• Deep learning models successfully identified subtle features, including gender, and Grad-CAM highlighted clinically relevant areas.

Conclusions:

• Image-based deep learning is a feasible approach for ECG diagnosis.
• This technology holds potential for developing clinically applicable tools for ECG analysis, especially where only images are accessible.
• Further research is needed to refine these models for widespread clinical adoption.