Leveraging ECG images for predicting ejection fraction using machine learning algorithms.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Predicting ejection fraction (EF) from electrocardiograms (ECGs) has significant clinical value.
• Current algorithms often require raw ECG signal data.
• Developing methods using readily available ECG images is crucial.

Purpose of the Study:

• To train and validate a neural network using ECG trace images.
• To determine the presence or absence of left ventricular dysfunction (LVD).
• To assess the feasibility of using ECG images for LVD screening.

Main Methods:

• A DenseNet121 model was trained on 12-lead ECG trace images.
• The model used paired ECG images and echocardiogram reports.
• The model was trained to identify EF <50% and externally validated.

Main Results:

• The study utilized 119,281 ECG-echocardiogram pairs.
• The model achieved high performance, with AUCs of 0.92 (internal) and 0.88 (external).
• The model accurately identified over 85% of cases with EF <50%.

Conclusions:

• ECG images, with simple pre-processing, serve as a reliable tool for LVD screening.
• This approach expands the utility of ECG-based algorithms in resource-limited areas.
• The findings support the use of AI on ECG images for cardiovascular assessment.