Background: Standard 12-lead electrocardiograms (ECGs) consist of twelve 1D signals, each capturing spatial cardiac electrical activity. This unique physiological format poses challenges for integration into both image- and signal-based deep learning models, often limiting the incorporation of spatial relationships between leads. Consequently, ECG research has not fully benefited from recent advances in computer vision. Meanwhile, Chagas disease—a tropical parasitic disease with vague and nonspecific clinical symptoms—remains difficult to diagnose in real-world clinical settings. We propose a novel approach that embeds ECG signals into 2D images, revealing underlying spatial information and enabling the use of pretrained image models for Chagas disease screening.
Methods: Based on the concept of Einthoven's triangle, six limb leads were interpolated across the coronal plane to construct a time-angle grayscale image. Horizontal and sagittal plane images were similarly generated using six precordial leads and a combination of aVF with (V1+V2)/2, respectively. These three anatomical plane images were assigned to RGB channels, resulting in a spatially embedded ECG image. A total of 365,061 ECG–label pairs from PhysioNet's CODE-15%, SaMi-Trop, and PTB-XL datasets were used to train an EfficientNetV2-Small model for binary classification of Chagas disease.
Results: The proposed method achieved a higher true positive rate in detecting Chagas cases within the top 5% of high-probability predictions, outperforming previous 1D signal-based approaches after just a few epochs of fine-tuning. In the unofficial phase of the challenge, our model achieved a score of 0.473 on the test set using 10 training epochs and a 5-fold ensemble.
Conclusion: This study demonstrates that embedding ECGs into spatially structured 2D images enables effective transfer learning with pretrained image models. The proposed embedding framework provides a scalable and generalizable strategy for integrating ECG analysis into state-of-the-art image models across a wide range of clinical tasks, including Chagas disease classification.