Beyond Echocardiography: Multimodal Deep Learning for Ejection Fraction Estimation from Standard ECG

Left Ventricular Dysfunction (LVD) is a critical cardiac abnormality strongly associated with deteriorating heart health. The Left Ventricular Ejection Fraction (LVEF) serves as a key indicator for assessing the degree of dysfunction. Early detection of LVD is vital to prevent the onset of heart failure (HF). Conventionally, LVEF is measured through echocardiography, a process that is often cumbersome, time-consuming, and resource-intensive. However, recent studies suggest that characteristic signatures of LVD can also be identified from electrocardiogram (ECG) signals. To address this, we propose a deep learning framework that leverages both temporal and frequency-domain representations of ECG signals to classify LVEF levels. The regression component of the model is guided by the classification probabilities, enhancing predictive performance. The framework directly processes raw 12-lead ECG signals. For frequency-domain analysis, Gramian Angular Field (GAF) representations are utilized, following dimensionality reduction via Principal Component Analysis (PCA). A late fusion strategy is employed to concatenate the feature representations derived from both modalities. The proposed architecture achieves a classification accuracy of 90.34\% and an F1-score of 93.34\%. For regression, the model attains a Lin’s Concordance Correlation Coefficient of 0.67 and a Pearson Correlation Coefficient of 0.69, both guided by the classification probabilities. These results demonstrate the model’s potential as an effective and efficient screening tool for LVD, significantly reducing the dependence on echocardiography for preliminary diagnosis.