Deep Learning for Early Chagas Disease Diagnosis: A Comparative Analysis of 12-Lead ECG and Derived VCG

Alejandro Pascual-Mellado1, Vicent Torres-Sastre2, Cristina Albert2, Alejandro Perez3, Raul Alos2, Elisa Ramirez4, Francisco Castells5, Jose Millet6
1Universidad Politecnica de Valencia, 2Universitat Politecnica de Valencia, 3ITACA Institute, 4Institute ITACA, Universitat Politecnica de Valencia, 5Universitat Politècnica de Valencia, 6BioITACA-UPV


Abstract

Chagas disease (ChD) is a chronic parasitic condition that can lead to severe cardiac complications. The use of electrocardiographic (ECG) analysis has emerged as a promising tool for early, non-invasive detection. This work, developed by the EPBandoleroLab team for the PhysioNet Challenge 2025, presents a deep learning approach for ChD classification using the CODE-15, SaMi-Trop, and PTB-XL databases. Our methodology explores the effectiveness of different signal representations, comparing the standard 12-lead ECG with the derived Vectorcardiogram (VCG). Furthermore, we address the significant class imbalance through a controlled sampling strategy. Our findings indicate that the model performs best when trained on the full 12-lead ECG representation with a moderately imbalanced dataset. This configuration achieved a Challenge Score of 0.259 in the official phase, placing our team in the top half of all competitors.