As part of the PhysioNet/Computing in Cardiology Challenge 2025, our team, AIChagas, trained a ResNet model to detect Chagas disease in the electrocardiogram (ECG) by managing label uncertainty from self-reporting data. We soft-labeled patients with multiple self-claims and negative samples with arrhythmias common in Chagas disease. For cases with multiple reports, we assigned soft probabilities according to their proportion of positive and total reports. In addition, we increased the soft probability of negative samples with arrhythmias commonly found in cases of Chagas cardiomyopathy. The soft probabilities of negative samples were capped to prevent relabeling. To avoid biases when training with multiple data sources, we employed several data processing and augmentation techniques. We introduced positive class weighting to mitigate the imbalance problem. Finally, we trained our model by combining binary cross-entropy (BCE) and margin ranking loss between positive and hard negative samples (20% of negative instances with the highest predicted scores) to enhance the model's ranking capability. In our internal evaluation, our approach achieved a challenge score of 0.402. In the hidden validation set, our model received a challenge score of 0.377 (ranked 30th out of 367 submissions).