Transfer Learning and Soft Labels Enable Robust ECG-Based Detection of Chagas Disease

Chagas disease (CD) is a tropical parasitic disease that often remains asymptomatic but can lead to serious long-term cardiac complications. This study describes our contribution to “Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025”, which focused on developing deep learning algorithms to detect CD using 12-lead electrocardiogram (ECG) data. We trained a convolutional neural network initialized with weights from ECGFounder, a foundation model pretrained on over 10 million ECGs. Model development used multinational datasets and applied preprocessing including filtering, resampling, and normalization of raw ECG signals. To address label noise, we applied soft labels to ECGs with features suggestive of asymptomatic CD, such as right bundle branch block and atrial fibrillation. The pretrained backbone was fine-tuned conservatively while the classification head was trained more aggressively using AdamW optimization and early stopping, before evaluation according to the challenge scoring metric. Our team, UMC Utrecht, achieved a challenge score of 0.192 on the hidden test set, resulting in a 27^th overall place. These findings underscore the potential of ECG-based tools for screening CD.