Myocardial scar and left ventricular ejection fraction classification for electrocardiography image using multi-task...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Myocardial scar (MS) and left ventricular ejection fraction (LVEF) are critical for cardiovascular assessment.
• Cardiac magnetic resonance (CMR) is the standard for MS and LVEF evaluation but is costly and inaccessible in many regions.
• Electrocardiograms (ECGs) offer a cost-effective alternative for cardiovascular diagnostics.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting MS and LVEF < 50% using 12-lead ECG 2D images.
• To compare the performance of 2D ECG image analysis with 1D signal analysis for these predictions.
• To assess the potential of ECG-based AI as a screening tool compared to CMR.

Main Methods:

• A multi-task deep learning framework was designed to analyze 14,052 12-lead ECG 2D images.
• Ground truth labels for MS and LVEF were obtained from CMR.
• The model's performance was evaluated using Area Under the Curve (AUC) metrics and compared against cardiologists' performance.

Main Results:

• The top-performing model achieved an AUC of 0.838 for MS prediction and 0.939 for LVEF < 50% classification.
• Model performance surpassed that of human cardiologists in these predictions.
• Analysis of 1D ECG signals showed inferior results compared to the 2D image-based approach.
• A prevalence-specific test dataset yielded an AUC of 0.812 for MS prediction.

Conclusions:

• Computer vision-based deep learning models can effectively classify MS and LVEF < 50% from ECG scan images.
• This AI-driven approach presents a cost-effective and accessible alternative to CMR for clinical screening.
• The 2D image analysis method demonstrates superior performance over 1D signal extraction for these cardiovascular parameters.