Prediction of reduced left ventricular ejection fraction using atrial fibrillation or flutter electrocardiograms: A...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Left ventricular ejection fraction (LVEF) evaluation is critical for managing patients with atrial fibrillation (AF) or atrial flutter (AFL).
• Predicting reduced LVEF (<50%) using AF/AFL electrocardiograms (ECGs) currently lacks sufficient evidence.
• Developing non-invasive methods to predict LVEF is essential for timely medical intervention.

Purpose of the Study:

• To investigate deep-learning approaches for predicting reduced LVEF (<50%) in patients with AF/AFL.
• To assess the efficacy of a customized convolutional neural network model (AFibEFNet) using ECGs and clinical data.
• To establish a reliable method for early identification of reduced LVEF in AF/AFL patients.

Main Methods:

• A cohort of 15,683 patients with AF/AFL and echocardiography data was analyzed.
• Deep-learning models, including the novel AFibEFNet, were trained using ECG signals, ECG features, and clinical information.
• Model performance was evaluated using a hold-out test dataset, five-fold cross-validation, and calibration plots.

Main Results:

• The AFibEFNet model demonstrated superior predictive performance compared to other models, achieving an AUROC of 0.798 with ECGs alone.
• Incorporating clinical features alongside ECG data significantly improved AFibEFNet's AUROC to 0.816 and AUPRC to 0.547.
• The model primarily focused on specific ECG components like the R-wave and QRS complex for prediction.

Conclusions:

• Machine learning, particularly deep learning models like AFibEFNet, can effectively predict reduced LVEF in patients with AF/AFL using standard 12-lead ECGs.
• The integration of clinical data enhances the predictive accuracy of these models.
• This approach offers a promising non-invasive tool for identifying patients at risk of reduced LVEF.