Ecg Image Based Predictions for Heart Care Using Machine Learning Based Framework

Abstract

Early detection of left ventricular dysfunction (LVD), especially in asymptomatic individuals, is critical for timely intervention and improved cardiovascular outcomes. However, widespread access to echocardiography remains limited, particularly in low-resource settings. In this study, we developed and externally validated a deep learning (DL) model that predicts left ventricular ejection fraction (LVEF) from 12-lead ECG trace images, eliminating the need for raw signal data. A total of 1,19,281 ECG-echocardiogram pairs from 1,04,697 patients formed the training and test datasets, while 24,319 pairs from a distinct cohort were used for external validation. The ECG trace plots were processed via multi-otsu thresholding to extract the region of interest and standardized using Z-score normalization. The model architecture was based on DenseNet121, trained with class-weighted focal binary cross-entropy loss to address data imbalance. On internal test data, the model achieved a receiver operating characteristic area under curve (ROCAUC) of 0.92 and precision-recall AUC (PRAUC) of 0.78 in identifying LVEF < 50%. External validation yielded comparable performance with ROCAUC and PRAUC of 0.88 and 0.74, respectively. Notably, the algorithm demonstrated 97% sensitivity in detecting severe LVD (EF ≤ 35%) and maintained robust performance across age, sex, and paced ECG subgroups. With a diagnostic odds ratio of 31.7 on test data and a high negative predictive value (NPV ~0.94), the model ensures low false negative rates—critical for triaging in high-volume clinical settings. This study highlights the feasibility of using ECG image-based DL models for LVD screening, especially in resource-constrained environments. The ability to extract LVEF-related features from trace images offers practical scalability in primary and tertiary healthcare centers and introduces a new paradigm in accessible, AI-powered cardiac diagnostics.