Background and Objective: Accurate detection of Chagas cardiomyopathy from electrocardiograms (ECGs) is critical for timely intervention, particularly in resource-limited settings. While conventional methods rely on handcrafted features or limited lead information, we explore the effectiveness of transformer-based deep learning models to detect Chagas disease from 12-lead ECGs across multiple populations and datasets.
Methods: We developed a hybrid CNN-transformer model to classify Chagas cardiomyopathy from 12-lead ECG signals. The model architecture begins with a series of convolutional layers to extract local temporal patterns, followed by max-pooling to downsample the feature maps. This is followed by stacked transformer encoder blocks that capture global dependencies across the signal through multi-head self-attention and feed-forward convolutional layers. The output is aggregated using global average pooling and passed through fully connected layers before producing a sigmoid-activated binary classification output. The model was trained using ECGs from the SaMi-Trop and CODE-15% datasets and evaluated on PTB-XL as an external dataset.
Results: The model showed generalizable performance across datasets, our team Mainchagas achievied an 0.691 challenge score.
Conclusion: This study highlights the potential of transformer-based architectures for robust ECG interpretation in the context of Chagas disease. The pipeline's performance across diverse datasets underscores its viability for deployment in scalable screening programs.